{ "generatedAt": "2026-11-03", "archs": [ { "name": "Blackwell", "short": "BLK", "years": "2024—2026", "tech": "FP4 · 双die · HBM3e", "first": "2024-03-01", "last": "2026-06-01", "count": 22 }, { "name": "Hopper", "short": "HOP", "years": "2022—2025", "tech": "Transformer Engine · HBM3", "first": "2022-03-01", "last": "2024-06-01", "count": 5 }, { "name": "Ada Lovelace", "short": "ADA", "years": "2022—2024", "tech": "DLSS 3 · 4N", "first": "2022-06-01", "last": "2024-06-01", "count": 21 }, { "name": "Ampere", "short": "AMP", "years": "2020—2022", "tech": "TF32 · MIG", "first": "2020-05-01", "last": "2022-06-01", "count": 25 }, { "name": "Turing", "short": "TUR", "years": "2018—2020", "tech": "RT Core · Tensor Core", "first": "2018-06-01", "last": "2019-11-01", "count": 17 }, { "name": "Volta", "short": "VOL", "years": "2017—2019", "tech": "首代 Tensor Core · HBM2", "first": "2017-06-01", "last": "2019-06-01", "count": 7 }, { "name": "Pascal", "short": "PAS", "years": "2016—2018", "tech": "FP16 · HBM2 · 16nm", "first": "2016-04-01", "last": "2017-11-01", "count": 17 }, { "name": "Maxwell", "short": "MAX", "years": "2014—2016", "tech": "能效跃迁", "first": "2014-06-01", "last": "2015-06-01", "count": 13 }, { "name": "Kepler", "short": "KEP", "years": "2012—2014", "tech": "GK110 · 动态并行", "first": "2012-06-01", "last": "2014-06-01", "count": 15 }, { "name": "Fermi", "short": "FER", "years": "2010—2011", "tech": "首代完整 GPGPU", "first": "2010-06-01", "last": "2011-06-01", "count": 12 } ], "products": [ { "id": "blk-b300blackwellultra", "name": "B300（Blackwell Ultra）", "arch": "Blackwell", "archShort": "BLK", "archYears": "2024—2026", "archTech": "FP4 · 双die · HBM3e", "tier": "datacenter", "category": "数据中心·AI训练/推理", "cuda": null, "vramGb": 288.0, "vramPerGpuGb": 288.0, "vramType": "HBM3e", "vramLabel": "288GB HBM3e", "fp32": 75.0, "date": "2025-03-01", "datePrecision": "month", "dateText": "2025年3月发布", "note": "双die封装；8TB/s；NVFP4（稀疏）约20 PFLOPS；对应GB300/GB200 NVL72超级芯片系统", "china": false, "brand": "NVIDIA", "source": "docx", "marketId": "b300_sxm", "msrpUsd": null, "priceCny": null, "regionAvail": "restricted", "perfAi": 220, "perfGaming": 0, "tdpW": 1400, "bandwidthGbps": 8000, "tensorTflops": 15000, "nvlink": true, "multiMax": 8, "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "blk-b200", "name": "B200", "arch": "Blackwell", "archShort": "BLK", "archYears": "2024—2026", "archTech": "FP4 · 双die · HBM3e", "tier": "datacenter", "category": "数据中心·AI训练/推理", "cuda": null, "vramGb": 192.0, "vramPerGpuGb": 192.0, "vramType": "HBM3e", "vramLabel": "192GB HBM3e", "fp32": 75.0, "date": "2024-03-01", "datePrecision": "month", "dateText": "2024年3月发布", "note": "8TB/s；NVFP4（稀疏）约18—20 PFLOPS", "china": false, "brand": "NVIDIA", "source": "docx", "marketId": "b200_sxm", "msrpUsd": 45000, "priceCny": 380000, "regionAvail": "restricted", "perfAi": 180, "perfGaming": 0, "tdpW": 1000, "bandwidthGbps": 8000, "tensorTflops": 9000, "nvlink": true, "multiMax": 8, "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "blk-b100", "name": "B100", "arch": "Blackwell", "archShort": "BLK", "archYears": "2024—2026", "archTech": "FP4 · 双die · HBM3e", "tier": "datacenter", "category": "数据中心·AI", "cuda": null, "vramGb": 180.0, "vramPerGpuGb": 180.0, "vramType": "HBM3e", "vramLabel": "180GB HBM3e", "fp32": null, "date": "2024-03-01", "datePrecision": "month", "dateText": "2024年3月发布", "note": "液冷为主、供货有限，主要面向云厂商", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "blk-rtxpro6000blackwell", "name": "RTX PRO 6000 Blackwell（工作站版）", "arch": "Blackwell", "archShort": "BLK", "archYears": "2024—2026", "archTech": "FP4 · 双die · HBM3e", "tier": "workstation", "category": "专业工作站", "cuda": 24064.0, "vramGb": 96.0, "vramPerGpuGb": 96.0, "vramType": "GDDR7", "vramLabel": "96GB GDDR7", "fp32": 126.0, "date": "2025-03-01", "datePrecision": "month", "dateText": "2025年3月", "note": "1792GB/s；AI TOPS 4000（FP4）；600W", "china": false, "brand": "NVIDIA", "source": "docx", "marketId": "rtx_pro6000_bw", "msrpUsd": 8435, "priceCny": 69999, "regionAvail": "in_stock", "perfAi": 110, "perfGaming": 82, "tdpW": 600, "bandwidthGbps": 1792, "tensorTflops": 4000, "nvlink": false, "multiMax": 4, "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "blk-rtxpro6000blackwell-2", "name": "RTX PRO 6000 Blackwell（服务器版）", "arch": "Blackwell", "archShort": "BLK", "archYears": "2024—2026", "archTech": "FP4 · 双die · HBM3e", "tier": "workstation", "category": "专业·服务器", "cuda": 24064.0, "vramGb": 96.0, "vramPerGpuGb": 96.0, "vramType": "GDDR7", "vramLabel": "96GB GDDR7", "fp32": 126.0, "date": "2025-03-01", "datePrecision": "month", "dateText": "2025年3月", "note": "被动散热；FP4（稀疏）4 PFLOPS；最高600W", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "blk-rtxpro6000blackwellmaxq", "name": "RTX PRO 6000 Blackwell（Max-Q工作站版）", "arch": "Blackwell", "archShort": "BLK", "archYears": "2024—2026", "archTech": "FP4 · 双die · HBM3e", "tier": "workstation", "category": "专业工作站", "cuda": 24064.0, "vramGb": 96.0, "vramPerGpuGb": 96.0, "vramType": "GDDR7", "vramLabel": "96GB GDDR7", "fp32": 110.0, "date": "2025-03-01", "datePrecision": "month", "dateText": "2025年3月", "note": "300W能效版；FP32约109.7", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "blk-rtx6000d", "name": "RTX 6000D", "arch": "Blackwell", "archShort": "BLK", "archYears": "2024—2026", "archTech": "FP4 · 双die · HBM3e", "tier": "workstation", "category": "专业·中国特供", "cuda": 19968.0, "vramGb": 84.0, "vramPerGpuGb": 84.0, "vramType": "GDDR7", "vramLabel": "84GB GDDR7", "fp32": 97.0, "date": "2026-06-01", "datePrecision": "year", "dateText": "2026年初", "note": "基于PRO 6000削减约17%核心；显存84GB/448bit，带宽1.57TB/s；被动散热", "china": true, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "blk-rtxpro5500blackwell", "name": "RTX PRO 5500 Blackwell（工作站版）", "arch": "Blackwell", "archShort": "BLK", "archYears": "2024—2026", "archTech": "FP4 · 双die · HBM3e", "tier": "workstation", "category": "专业·机架工作站", "cuda": null, "vramGb": 84.0, "vramPerGpuGb": 84.0, "vramType": "GDDR7", "vramLabel": "84GB GDDR7", "fp32": null, "date": "2026-06-01", "datePrecision": "year", "dateText": "2026年发布（待上市）", "note": "1398GB/s；最高600W；支持MIG（2实例）", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "blk-rtxpro5000blackwell", "name": "RTX PRO 5000 Blackwell", "arch": "Blackwell", "archShort": "BLK", "archYears": "2024—2026", "archTech": "FP4 · 双die · HBM3e", "tier": "workstation", "category": "专业工作站", "cuda": 14080.0, "vramGb": 72.0, "vramPerGpuGb": 72.0, "vramType": "GDDR7", "vramLabel": "48GB GDDR7（另有72GB版）", "fp32": null, "date": "2025-03-01", "datePrecision": "month", "dateText": "2025年3月", "note": "300W；1344GB/s", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "blk-rtxpro4500blackwell", "name": "RTX PRO 4500 Blackwell（工作站版）", "arch": "Blackwell", "archShort": "BLK", "archYears": "2024—2026", "archTech": "FP4 · 双die · HBM3e", "tier": "workstation", "category": "专业工作站", "cuda": null, "vramGb": 32.0, "vramPerGpuGb": 32.0, "vramType": "GDDR7", "vramLabel": "32GB GDDR7", "fp32": null, "date": "2025-03-01", "datePrecision": "month", "dateText": "2025年3月", "note": "200W", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "blk-rtxpro4000blackwell", "name": "RTX PRO 4000 Blackwell", "arch": "Blackwell", "archShort": "BLK", "archYears": "2024—2026", "archTech": "FP4 · 双die · HBM3e", "tier": "workstation", "category": "专业·单槽", "cuda": null, "vramGb": 24.0, "vramPerGpuGb": 24.0, "vramType": "GDDR7", "vramLabel": "24GB GDDR7", "fp32": null, "date": "2025-07-01", "datePrecision": "month", "dateText": "2025年7月", "note": "145W", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "blk-rtxpro4000blackwellsff", "name": "RTX PRO 4000 Blackwell SFF", "arch": "Blackwell", "archShort": "BLK", "archYears": "2024—2026", "archTech": "FP4 · 双die · HBM3e", "tier": "workstation", "category": "专业·小尺寸", "cuda": null, "vramGb": 24.0, "vramPerGpuGb": 24.0, "vramType": "GDDR7", "vramLabel": "24GB GDDR7", "fp32": null, "date": "2025-06-01", "datePrecision": "year", "dateText": "2025年下半年", "note": "70W", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "blk-rtxpro2000blackwell", "name": "RTX PRO 2000 Blackwell", "arch": "Blackwell", "archShort": "BLK", "archYears": "2024—2026", "archTech": "FP4 · 双die · HBM3e", "tier": "workstation", "category": "专业·入门", "cuda": null, "vramGb": 16.0, "vramPerGpuGb": 16.0, "vramType": "GDDR7", "vramLabel": "16GB GDDR7", "fp32": null, "date": "2025-06-01", "datePrecision": "year", "dateText": "2025年下半年", "note": "70W", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "blk-rtx5090", "name": "RTX 5090", "arch": "Blackwell", "archShort": "BLK", "archYears": "2024—2026", "archTech": "FP4 · 双die · HBM3e", "tier": "consumer", "category": "消费级", "cuda": 21760.0, "vramGb": 32.0, "vramPerGpuGb": 32.0, "vramType": "GDDR7", "vramLabel": "32GB GDDR7", "fp32": 105.0, "date": "2025-01-01", "datePrecision": "month", "dateText": "2025年1月", "note": "1792GB/s；AI TOPS 3352（FP4）", "china": false, "brand": "NVIDIA", "source": "docx", "marketId": "rtx5090", "msrpUsd": 1999, "priceCny": 19999, "regionAvail": "limited", "perfAi": 98, "perfGaming": 100, "tdpW": 575, "bandwidthGbps": 1792, "tensorTflops": 1676, "nvlink": false, "multiMax": 1, "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "blk-rtx5090d", "name": "RTX 5090D", "arch": "Blackwell", "archShort": "BLK", "archYears": "2024—2026", "archTech": "FP4 · 双die · HBM3e", "tier": "consumer", "category": "消费级·中国特供", "cuda": 21760.0, "vramGb": 32.0, "vramPerGpuGb": 32.0, "vramType": "GDDR7", "vramLabel": "32GB GDDR7", "fp32": 105.0, "date": "2025-01-01", "datePrecision": "month", "dateText": "2025年1月", "note": "去除NVLink；AI TOPS 2375（官方口径）", "china": true, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "blk-rtx5090dv2", "name": "RTX 5090D V2", "arch": "Blackwell", "archShort": "BLK", "archYears": "2024—2026", "archTech": "FP4 · 双die · HBM3e", "tier": "consumer", "category": "消费级·中国特供", "cuda": 21760.0, "vramGb": 24.0, "vramPerGpuGb": 24.0, "vramType": "GDDR7", "vramLabel": "24GB GDDR7", "fp32": 105.0, "date": "2025-10-01", "datePrecision": "month", "dateText": "2025年10月", "note": "5090D新版：显存降至24GB/384bit，带宽1344GB/s", "china": true, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "blk-rtx5080", "name": "RTX 5080", "arch": "Blackwell", "archShort": "BLK", "archYears": "2024—2026", "archTech": "FP4 · 双die · HBM3e", "tier": "consumer", "category": "消费级", "cuda": 10752.0, "vramGb": 16.0, "vramPerGpuGb": 16.0, "vramType": "GDDR7", "vramLabel": "16GB GDDR7", "fp32": 56.0, "date": "2025-01-01", "datePrecision": "month", "dateText": "2025年1月", "note": "256bit", "china": false, "brand": "NVIDIA", "source": "docx", "marketId": "rtx5080", "msrpUsd": 999, "priceCny": 9999, "regionAvail": "in_stock", "perfAi": 72, "perfGaming": 78, "tdpW": 360, "bandwidthGbps": 960, "tensorTflops": 900, "nvlink": false, "multiMax": 1, "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "blk-rtx5070ti", "name": "RTX 5070 Ti", "arch": "Blackwell", "archShort": "BLK", "archYears": "2024—2026", "archTech": "FP4 · 双die · HBM3e", "tier": "consumer", "category": "消费级", "cuda": 8960.0, "vramGb": 16.0, "vramPerGpuGb": 16.0, "vramType": "GDDR7", "vramLabel": "16GB GDDR7", "fp32": 44.0, "date": "2025-02-01", "datePrecision": "month", "dateText": "2025年2月", "note": "256bit", "china": false, "brand": "NVIDIA", "source": "docx", "marketId": "rtx5070ti", "msrpUsd": 749, "priceCny": 6499, "regionAvail": "in_stock", "perfAi": 58, "perfGaming": 62, "tdpW": 300, "bandwidthGbps": 896, "tensorTflops": 706, "nvlink": false, "multiMax": 1, "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "blk-rtx5070", "name": "RTX 5070", "arch": "Blackwell", "archShort": "BLK", "archYears": "2024—2026", "archTech": "FP4 · 双die · HBM3e", "tier": "consumer", "category": "消费级", "cuda": 6144.0, "vramGb": 12.0, "vramPerGpuGb": 12.0, "vramType": "GDDR7", "vramLabel": "12GB GDDR7", "fp32": 31.0, "date": "2025-03-01", "datePrecision": "month", "dateText": "2025年3月", "note": "192bit", "china": false, "brand": "NVIDIA", "source": "docx", "marketId": "rtx5070", "msrpUsd": 549, "priceCny": 4599, "regionAvail": "in_stock", "perfAi": 52, "perfGaming": 56, "tdpW": 250, "bandwidthGbps": 672, "tensorTflops": 494, "nvlink": false, "multiMax": 1, "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "blk-rtx5060ti", "name": "RTX 5060 Ti", "arch": "Blackwell", "archShort": "BLK", "archYears": "2024—2026", "archTech": "FP4 · 双die · HBM3e", "tier": "consumer", "category": "消费级", "cuda": 4608.0, "vramGb": 16.0, "vramPerGpuGb": 16.0, "vramType": "GDDR7", "vramLabel": "16GB/8GB GDDR7", "fp32": 24.0, "date": "2025-04-01", "datePrecision": "month", "dateText": "2025年4月", "note": "128bit", "china": false, "brand": "NVIDIA", "source": "docx", "marketId": "rtx5060ti_16", "msrpUsd": 429, "priceCny": 3599, "regionAvail": "in_stock", "perfAi": 38, "perfGaming": 42, "tdpW": 180, "bandwidthGbps": 448, "tensorTflops": 384, "nvlink": false, "multiMax": 1, "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "blk-rtx5060", "name": "RTX 5060", "arch": "Blackwell", "archShort": "BLK", "archYears": "2024—2026", "archTech": "FP4 · 双die · HBM3e", "tier": "consumer", "category": "消费级", "cuda": 3840.0, "vramGb": 8.0, "vramPerGpuGb": 8.0, "vramType": "GDDR7", "vramLabel": "8GB GDDR7", "fp32": 19.0, "date": "2025-05-01", "datePrecision": "month", "dateText": "2025年5月", "note": "128bit", "china": false, "brand": "NVIDIA", "source": "docx", "marketId": "rtx5060", "msrpUsd": 299, "priceCny": 2499, "regionAvail": "in_stock", "perfAi": 30, "perfGaming": 34, "tdpW": 145, "bandwidthGbps": 448, "tensorTflops": 309, "nvlink": false, "multiMax": 1, "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "blk-rtx5050", "name": "RTX 5050", "arch": "Blackwell", "archShort": "BLK", "archYears": "2024—2026", "archTech": "FP4 · 双die · HBM3e", "tier": "consumer", "category": "消费级", "cuda": 2560.0, "vramGb": 8.0, "vramPerGpuGb": 8.0, "vramType": "GDDR6", "vramLabel": "8GB GDDR6", "fp32": 13.0, "date": "2025-07-01", "datePrecision": "month", "dateText": "2025年7月", "note": "128bit；320GB/s", "china": false, "brand": "NVIDIA", "source": "docx", "marketId": "rtx5050", "msrpUsd": 249, "priceCny": 2099, "regionAvail": "in_stock", "perfAi": 20, "perfGaming": 26, "tdpW": 130, "bandwidthGbps": 320, "tensorTflops": 241, "nvlink": false, "multiMax": 1, "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "hop-h200sxm", "name": "H200 SXM", "arch": "Hopper", "archShort": "HOP", "archYears": "2022—2025", "archTech": "Transformer Engine · HBM3", "tier": "datacenter", "category": "数据中心·AI/HPC", "cuda": 16896.0, "vramGb": 141.0, "vramPerGpuGb": 141.0, "vramType": "HBM3e", "vramLabel": "141GB HBM3e", "fp32": 67.0, "date": "2023-11-01", "datePrecision": "month", "dateText": "2023年11月发布", "note": "4.8TB/s；FP8张量3958（稀疏）；另有H200 NVL", "china": false, "brand": "NVIDIA", "source": "docx", "marketId": "h200_sxm", "msrpUsd": 35000, "priceCny": 300000, "regionAvail": "restricted", "perfAi": 120, "perfGaming": 0, "tdpW": 700, "bandwidthGbps": 4800, "tensorTflops": 3958, "nvlink": true, "multiMax": 8, "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "hop-h100sxm", "name": "H100 SXM", "arch": "Hopper", "archShort": "HOP", "archYears": "2022—2025", "archTech": "Transformer Engine · HBM3", "tier": "datacenter", "category": "数据中心·AI/HPC", "cuda": 16896.0, "vramGb": 80.0, "vramPerGpuGb": 80.0, "vramType": "HBM3", "vramLabel": "80GB HBM3", "fp32": 67.0, "date": "2022-03-01", "datePrecision": "month", "dateText": "2022年3月", "note": "3.35TB/s；FP8张量3958（稀疏）；另有GH200超级芯片（Grace+H100）", "china": false, "brand": "NVIDIA", "source": "docx", "marketId": "h100_sxm", "msrpUsd": 30000, "priceCny": 260000, "regionAvail": "restricted", "perfAi": 100, "perfGaming": 0, "tdpW": 700, "bandwidthGbps": 3350, "tensorTflops": 3958, "nvlink": true, "multiMax": 8, "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "hop-h100pcie", "name": "H100 PCIe", "arch": "Hopper", "archShort": "HOP", "archYears": "2022—2025", "archTech": "Transformer Engine · HBM3", "tier": "datacenter", "category": "数据中心", "cuda": null, "vramGb": 80.0, "vramPerGpuGb": 80.0, "vramType": "HBM3", "vramLabel": "80GB HBM3", "fp32": null, "date": "2022-06-01", "datePrecision": "year", "dateText": "2022年", "note": "PCIe版，规格略低于SXM版", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "hop-h800", "name": "H800", "arch": "Hopper", "archShort": "HOP", "archYears": "2022—2025", "archTech": "Transformer Engine · HBM3", "tier": "datacenter", "category": "数据中心·中国特供", "cuda": 16896.0, "vramGb": 80.0, "vramPerGpuGb": 80.0, "vramType": "HBM3", "vramLabel": "80GB HBM3", "fp32": 67.0, "date": "2022-06-01", "datePrecision": "year", "dateText": "2022年", "note": "中国版H100；NVLink降至400GB/s", "china": true, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "hop-h20", "name": "H20", "arch": "Hopper", "archShort": "HOP", "archYears": "2022—2025", "archTech": "Transformer Engine · HBM3", "tier": "datacenter", "category": "数据中心·中国特供", "cuda": null, "vramGb": 96.0, "vramPerGpuGb": 96.0, "vramType": "HBM3", "vramLabel": "96GB HBM3", "fp32": null, "date": "2024-06-01", "datePrecision": "year", "dateText": "2024年", "note": "中国版；4TB/s；FP16（稠密）148 TFLOPS；NVLink 900GB/s", "china": true, "brand": "NVIDIA", "source": "docx", "fp16Note": 148.0, "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "ada-l40s", "name": "L40S", "arch": "Ada Lovelace", "archShort": "ADA", "archYears": "2022—2024", "archTech": "DLSS 3 · 4N", "tier": "datacenter", "category": "数据中心·AI/图形", "cuda": 18176.0, "vramGb": 48.0, "vramPerGpuGb": 48.0, "vramType": "GDDR6", "vramLabel": "48GB GDDR6", "fp32": 91.6, "date": "2023-06-01", "datePrecision": "year", "dateText": "2023年", "note": "864GB/s；FP16张量733（稀疏）；AI推理强化版", "china": false, "brand": "NVIDIA", "source": "docx", "marketId": "l40s", "msrpUsd": 9000, "priceCny": 72000, "regionAvail": "in_stock", "perfAi": 72, "perfGaming": 0, "tdpW": 350, "bandwidthGbps": 864, "tensorTflops": 733, "nvlink": false, "multiMax": 8, "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "ada-l40", "name": "L40", "arch": "Ada Lovelace", "archShort": "ADA", "archYears": "2022—2024", "archTech": "DLSS 3 · 4N", "tier": "datacenter", "category": "数据中心·AI/图形", "cuda": 18176.0, "vramGb": 48.0, "vramPerGpuGb": 48.0, "vramType": "GDDR6", "vramLabel": "48GB GDDR6", "fp32": 90.5, "date": "2022-06-01", "datePrecision": "year", "dateText": "2022年", "note": "864GB/s；FP16张量约362（稀疏）", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "ada-l20", "name": "L20", "arch": "Ada Lovelace", "archShort": "ADA", "archYears": "2022—2024", "archTech": "DLSS 3 · 4N", "tier": "datacenter", "category": "数据中心·中国特供", "cuda": 10752.0, "vramGb": 48.0, "vramPerGpuGb": 48.0, "vramType": "GDDR6", "vramLabel": "48GB GDDR6", "fp32": null, "date": "2023-06-01", "datePrecision": "year", "dateText": "2023年", "note": "864GB/s；275W", "china": true, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "ada-l4", "name": "L4", "arch": "Ada Lovelace", "archShort": "ADA", "archYears": "2022—2024", "archTech": "DLSS 3 · 4N", "tier": "datacenter", "category": "数据中心·推理/边缘", "cuda": null, "vramGb": 24.0, "vramPerGpuGb": 24.0, "vramType": "GDDR6", "vramLabel": "24GB GDDR6", "fp32": 30.3, "date": "2023-06-01", "datePrecision": "year", "dateText": "2023年", "note": "FP8张量485（稀疏）；72W", "china": false, "brand": "NVIDIA", "source": "docx", "marketId": "l4", "msrpUsd": 2500, "priceCny": 19999, "regionAvail": "in_stock", "perfAi": 42, "perfGaming": 0, "tdpW": 72, "bandwidthGbps": 300, "tensorTflops": 485, "nvlink": false, "multiMax": 8, "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "ada-l2", "name": "L2", "arch": "Ada Lovelace", "archShort": "ADA", "archYears": "2022—2024", "archTech": "DLSS 3 · 4N", "tier": "datacenter", "category": "数据中心·入门", "cuda": null, "vramGb": 24.0, "vramPerGpuGb": 24.0, "vramType": "GDDR6", "vramLabel": "24GB GDDR6", "fp32": null, "date": "2023-06-01", "datePrecision": "year", "dateText": "2023年", "note": "60W", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "ada-rtx6000ada", "name": "RTX 6000 Ada", "arch": "Ada Lovelace", "archShort": "ADA", "archYears": "2022—2024", "archTech": "DLSS 3 · 4N", "tier": "workstation", "category": "专业工作站", "cuda": 18176.0, "vramGb": 48.0, "vramPerGpuGb": 48.0, "vramType": "GDDR6", "vramLabel": "48GB GDDR6", "fp32": 91.1, "date": "2022-12-01", "datePrecision": "month", "dateText": "2022年12月", "note": "960GB/s；300W", "china": false, "brand": "NVIDIA", "source": "docx", "marketId": "rtx6000ada", "msrpUsd": 6799, "priceCny": 54999, "regionAvail": "in_stock", "perfAi": 78, "perfGaming": 70, "tdpW": 300, "bandwidthGbps": 960, "tensorTflops": 1457, "nvlink": true, "multiMax": 4, "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "ada-rtx5000ada", "name": "RTX 5000 Ada", "arch": "Ada Lovelace", "archShort": "ADA", "archYears": "2022—2024", "archTech": "DLSS 3 · 4N", "tier": "workstation", "category": "专业工作站", "cuda": 12800.0, "vramGb": 32.0, "vramPerGpuGb": 32.0, "vramType": "GDDR6", "vramLabel": "32GB GDDR6", "fp32": 65.0, "date": "2023-06-01", "datePrecision": "year", "dateText": "2023年", "note": "250W", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "ada-rtx4500ada", "name": "RTX 4500 Ada", "arch": "Ada Lovelace", "archShort": "ADA", "archYears": "2022—2024", "archTech": "DLSS 3 · 4N", "tier": "workstation", "category": "专业工作站", "cuda": 7680.0, "vramGb": 24.0, "vramPerGpuGb": 24.0, "vramType": "GDDR6", "vramLabel": "24GB GDDR6", "fp32": 39.0, "date": "2023-06-01", "datePrecision": "year", "dateText": "2023年", "note": "210W", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "ada-rtx4000ada", "name": "RTX 4000 Ada", "arch": "Ada Lovelace", "archShort": "ADA", "archYears": "2022—2024", "archTech": "DLSS 3 · 4N", "tier": "workstation", "category": "专业工作站", "cuda": 6144.0, "vramGb": 20.0, "vramPerGpuGb": 20.0, "vramType": "GDDR6", "vramLabel": "20GB GDDR6", "fp32": 27.0, "date": "2023-06-01", "datePrecision": "year", "dateText": "2023年", "note": "130W", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "ada-rtx4000sffada", "name": "RTX 4000 SFF Ada", "arch": "Ada Lovelace", "archShort": "ADA", "archYears": "2022—2024", "archTech": "DLSS 3 · 4N", "tier": "workstation", "category": "专业·小尺寸", "cuda": 6144.0, "vramGb": 20.0, "vramPerGpuGb": 20.0, "vramType": "GDDR6", "vramLabel": "20GB GDDR6", "fp32": 27.0, "date": "2023-06-01", "datePrecision": "year", "dateText": "2023年", "note": "70W", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "ada-rtx2000ada", "name": "RTX 2000 Ada", "arch": "Ada Lovelace", "archShort": "ADA", "archYears": "2022—2024", "archTech": "DLSS 3 · 4N", "tier": "workstation", "category": "专业·入门", "cuda": 2816.0, "vramGb": 16.0, "vramPerGpuGb": 16.0, "vramType": "GDDR6", "vramLabel": "16GB GDDR6", "fp32": null, "date": "2024-06-01", "datePrecision": "year", "dateText": "2024年", "note": "70W", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "ada-rtx4090", "name": "RTX 4090", "arch": "Ada Lovelace", "archShort": "ADA", "archYears": "2022—2024", "archTech": "DLSS 3 · 4N", "tier": "consumer", "category": "消费级", "cuda": 16384.0, "vramGb": 24.0, "vramPerGpuGb": 24.0, "vramType": "GDDR6X", "vramLabel": "24GB GDDR6X", "fp32": 82.6, "date": "2022-10-01", "datePrecision": "month", "dateText": "2022年10月", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "marketId": "rtx4090", "msrpUsd": 1599, "priceCny": 14999, "regionAvail": "limited", "perfAi": 88, "perfGaming": 86, "tdpW": 450, "bandwidthGbps": 1008, "tensorTflops": 1321, "nvlink": false, "multiMax": 2, "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "ada-rtx4090d", "name": "RTX 4090D", "arch": "Ada Lovelace", "archShort": "ADA", "archYears": "2022—2024", "archTech": "DLSS 3 · 4N", "tier": "consumer", "category": "消费级·中国特供", "cuda": 14592.0, "vramGb": 24.0, "vramPerGpuGb": 24.0, "vramType": "GDDR6X", "vramLabel": "24GB GDDR6X", "fp32": 74.0, "date": "2023-12-01", "datePrecision": "month", "dateText": "2023年12月", "note": "中国特供", "china": true, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "ada-rtx4080super", "name": "RTX 4080 Super", "arch": "Ada Lovelace", "archShort": "ADA", "archYears": "2022—2024", "archTech": "DLSS 3 · 4N", "tier": "consumer", "category": "消费级", "cuda": 10240.0, "vramGb": 16.0, "vramPerGpuGb": 16.0, "vramType": "GDDR6X", "vramLabel": "16GB GDDR6X", "fp32": 52.0, "date": "2024-01-01", "datePrecision": "month", "dateText": "2024年1月", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "marketId": "rtx4080s", "msrpUsd": 999, "priceCny": 8299, "regionAvail": "in_stock", "perfAi": 62, "perfGaming": 68, "tdpW": 320, "bandwidthGbps": 736, "tensorTflops": 836, "nvlink": false, "multiMax": 1, "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "ada-rtx4080", "name": "RTX 4080", "arch": "Ada Lovelace", "archShort": "ADA", "archYears": "2022—2024", "archTech": "DLSS 3 · 4N", "tier": "consumer", "category": "消费级", "cuda": 9728.0, "vramGb": 16.0, "vramPerGpuGb": 16.0, "vramType": "GDDR6X", "vramLabel": "16GB GDDR6X", "fp32": 49.0, "date": "2022-11-01", "datePrecision": "month", "dateText": "2022年11月", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "ada-rtx4070tisuper", "name": "RTX 4070 Ti Super", "arch": "Ada Lovelace", "archShort": "ADA", "archYears": "2022—2024", "archTech": "DLSS 3 · 4N", "tier": "consumer", "category": "消费级", "cuda": 8448.0, "vramGb": 16.0, "vramPerGpuGb": 16.0, "vramType": "GDDR6X", "vramLabel": "16GB GDDR6X", "fp32": 44.0, "date": "2024-01-01", "datePrecision": "month", "dateText": "2024年1月", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "ada-rtx4070ti", "name": "RTX 4070 Ti", "arch": "Ada Lovelace", "archShort": "ADA", "archYears": "2022—2024", "archTech": "DLSS 3 · 4N", "tier": "consumer", "category": "消费级", "cuda": 7680.0, "vramGb": 12.0, "vramPerGpuGb": 12.0, "vramType": "GDDR6X", "vramLabel": "12GB GDDR6X", "fp32": 40.0, "date": "2023-01-01", "datePrecision": "month", "dateText": "2023年1月", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "ada-rtx4070super", "name": "RTX 4070 Super", "arch": "Ada Lovelace", "archShort": "ADA", "archYears": "2022—2024", "archTech": "DLSS 3 · 4N", "tier": "consumer", "category": "消费级", "cuda": 7168.0, "vramGb": 12.0, "vramPerGpuGb": 12.0, "vramType": "GDDR6X", "vramLabel": "12GB GDDR6X", "fp32": 35.0, "date": "2024-01-01", "datePrecision": "month", "dateText": "2024年1月", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "marketId": "rtx4070s", "msrpUsd": 599, "priceCny": 4899, "regionAvail": "in_stock", "perfAi": 42, "perfGaming": 48, "tdpW": 220, "bandwidthGbps": 504, "tensorTflops": 568, "nvlink": false, "multiMax": 1, "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "ada-rtx4070", "name": "RTX 4070", "arch": "Ada Lovelace", "archShort": "ADA", "archYears": "2022—2024", "archTech": "DLSS 3 · 4N", "tier": "consumer", "category": "消费级", "cuda": 5888.0, "vramGb": 12.0, "vramPerGpuGb": 12.0, "vramType": "GDDR6X", "vramLabel": "12GB GDDR6X", "fp32": 29.0, "date": "2023-04-01", "datePrecision": "month", "dateText": "2023年4月", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "ada-rtx4060ti", "name": "RTX 4060 Ti", "arch": "Ada Lovelace", "archShort": "ADA", "archYears": "2022—2024", "archTech": "DLSS 3 · 4N", "tier": "consumer", "category": "消费级", "cuda": 4352.0, "vramGb": 16.0, "vramPerGpuGb": 16.0, "vramType": "GDDR6", "vramLabel": "8GB/16GB GDDR6", "fp32": 22.0, "date": "2023-05-01", "datePrecision": "month", "dateText": "2023年5月", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "marketId": "rtx4060ti_16", "msrpUsd": 499, "priceCny": 3299, "regionAvail": "in_stock", "perfAi": 28, "perfGaming": 32, "tdpW": 165, "bandwidthGbps": 288, "tensorTflops": 353, "nvlink": false, "multiMax": 1, "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "ada-rtx4060", "name": "RTX 4060", "arch": "Ada Lovelace", "archShort": "ADA", "archYears": "2022—2024", "archTech": "DLSS 3 · 4N", "tier": "consumer", "category": "消费级", "cuda": 3072.0, "vramGb": 8.0, "vramPerGpuGb": 8.0, "vramType": "GDDR6", "vramLabel": "8GB GDDR6", "fp32": 15.0, "date": "2023-06-01", "datePrecision": "month", "dateText": "2023年6月", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "marketId": "rtx4060", "msrpUsd": 299, "priceCny": 2399, "regionAvail": "in_stock", "perfAi": 22, "perfGaming": 28, "tdpW": 115, "bandwidthGbps": 272, "tensorTflops": 242, "nvlink": false, "multiMax": 1, "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "amp-a100sxm80gb", "name": "A100 SXM（80GB）", "arch": "Ampere", "archShort": "AMP", "archYears": "2020—2022", "archTech": "TF32 · MIG", "tier": "datacenter", "category": "数据中心·AI/HPC", "cuda": 6912.0, "vramGb": 80.0, "vramPerGpuGb": 80.0, "vramType": "HBM2e", "vramLabel": "80GB HBM2e", "fp32": 19.5, "date": "2020-05-01", "datePrecision": "month", "dateText": "2020年5月", "note": "2039GB/s；FP16张量（稠密）312 TFLOPS", "china": false, "brand": "NVIDIA", "source": "docx", "marketId": "a100_80", "msrpUsd": 15000, "priceCny": 95000, "regionAvail": "in_stock", "perfAi": 68, "perfGaming": 0, "tdpW": 400, "bandwidthGbps": 2039, "tensorTflops": 624, "nvlink": true, "multiMax": 8, "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "amp-a100pcie80gb40gb", "name": "A100 PCIe（80GB/40GB）", "arch": "Ampere", "archShort": "AMP", "archYears": "2020—2022", "archTech": "TF32 · MIG", "tier": "datacenter", "category": "数据中心", "cuda": 6912.0, "vramGb": 80.0, "vramPerGpuGb": 80.0, "vramType": "HBM2e", "vramLabel": "80GB/40GB HBM2e", "fp32": 19.5, "date": "2020-06-01", "datePrecision": "year", "dateText": "2020年", "note": "PCIe版", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "amp-a800", "name": "A800", "arch": "Ampere", "archShort": "AMP", "archYears": "2020—2022", "archTech": "TF32 · MIG", "tier": "datacenter", "category": "数据中心·中国特供", "cuda": 6912.0, "vramGb": 80.0, "vramPerGpuGb": 80.0, "vramType": "HBM2e", "vramLabel": "80GB HBM2e", "fp32": 19.5, "date": "2022-06-01", "datePrecision": "year", "dateText": "2022年", "note": "中国版A100；NVLink降至400GB/s", "china": true, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "amp-a40", "name": "A40", "arch": "Ampere", "archShort": "AMP", "archYears": "2020—2022", "archTech": "TF32 · MIG", "tier": "datacenter", "category": "数据中心·AI/图形", "cuda": 10752.0, "vramGb": 48.0, "vramPerGpuGb": 48.0, "vramType": "GDDR6", "vramLabel": "48GB GDDR6", "fp32": 37.4, "date": "2020-06-01", "datePrecision": "year", "dateText": "2020年", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "amp-a10", "name": "A10", "arch": "Ampere", "archShort": "AMP", "archYears": "2020—2022", "archTech": "TF32 · MIG", "tier": "datacenter", "category": "数据中心·推理", "cuda": 9216.0, "vramGb": 24.0, "vramPerGpuGb": 24.0, "vramType": "GDDR6", "vramLabel": "24GB GDDR6", "fp32": 31.2, "date": "2021-06-01", "datePrecision": "year", "dateText": "2021年", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "amp-a30", "name": "A30", "arch": "Ampere", "archShort": "AMP", "archYears": "2020—2022", "archTech": "TF32 · MIG", "tier": "datacenter", "category": "数据中心·推理", "cuda": null, "vramGb": 24.0, "vramPerGpuGb": 24.0, "vramType": "HBM2", "vramLabel": "24GB HBM2", "fp32": 10.3, "date": "2020-06-01", "datePrecision": "year", "dateText": "2020年", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "amp-a16", "name": "A16", "arch": "Ampere", "archShort": "AMP", "archYears": "2020—2022", "archTech": "TF32 · MIG", "tier": "datacenter", "category": "数据中心·推理", "cuda": null, "vramGb": 64.0, "vramPerGpuGb": 16.0, "vramType": "GDDR6", "vramLabel": "4×16GB GDDR6（共64GB）", "fp32": null, "date": "2021-06-01", "datePrecision": "year", "dateText": "2021年", "note": "四路GPU合一板卡", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "amp-a2", "name": "A2", "arch": "Ampere", "archShort": "AMP", "archYears": "2020—2022", "archTech": "TF32 · MIG", "tier": "datacenter", "category": "数据中心·边缘", "cuda": null, "vramGb": 16.0, "vramPerGpuGb": 16.0, "vramType": "GDDR6", "vramLabel": "16GB GDDR6", "fp32": null, "date": "2021-06-01", "datePrecision": "year", "dateText": "2021年", "note": "60W", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "amp-rtxa6000", "name": "RTX A6000", "arch": "Ampere", "archShort": "AMP", "archYears": "2020—2022", "archTech": "TF32 · MIG", "tier": "workstation", "category": "专业工作站", "cuda": 10752.0, "vramGb": 48.0, "vramPerGpuGb": 48.0, "vramType": "GDDR6", "vramLabel": "48GB GDDR6", "fp32": 38.7, "date": "2020-06-01", "datePrecision": "year", "dateText": "2020年", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "amp-rtxa5500", "name": "RTX A5500", "arch": "Ampere", "archShort": "AMP", "archYears": "2020—2022", "archTech": "TF32 · MIG", "tier": "workstation", "category": "专业工作站", "cuda": 10240.0, "vramGb": 24.0, "vramPerGpuGb": 24.0, "vramType": "GDDR6", "vramLabel": "24GB GDDR6", "fp32": null, "date": "2021-06-01", "datePrecision": "year", "dateText": "2021年", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "amp-rtxa5000", "name": "RTX A5000", "arch": "Ampere", "archShort": "AMP", "archYears": "2020—2022", "archTech": "TF32 · MIG", "tier": "workstation", "category": "专业工作站", "cuda": 8192.0, "vramGb": 24.0, "vramPerGpuGb": 24.0, "vramType": "GDDR6", "vramLabel": "24GB GDDR6", "fp32": 27.8, "date": "2021-06-01", "datePrecision": "year", "dateText": "2021年", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "amp-rtxa4500", "name": "RTX A4500", "arch": "Ampere", "archShort": "AMP", "archYears": "2020—2022", "archTech": "TF32 · MIG", "tier": "workstation", "category": "专业工作站", "cuda": 7168.0, "vramGb": 20.0, "vramPerGpuGb": 20.0, "vramType": "GDDR6", "vramLabel": "20GB GDDR6", "fp32": null, "date": "2021-06-01", "datePrecision": "year", "dateText": "2021年", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "amp-rtxa4000", "name": "RTX A4000", "arch": "Ampere", "archShort": "AMP", "archYears": "2020—2022", "archTech": "TF32 · MIG", "tier": "workstation", "category": "专业工作站", "cuda": 6144.0, "vramGb": 16.0, "vramPerGpuGb": 16.0, "vramType": "GDDR6", "vramLabel": "16GB GDDR6", "fp32": 19.2, "date": "2021-06-01", "datePrecision": "year", "dateText": "2021年", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "amp-rtxa200012gb6gb", "name": "RTX A2000（12GB/6GB）", "arch": "Ampere", "archShort": "AMP", "archYears": "2020—2022", "archTech": "TF32 · MIG", "tier": "workstation", "category": "专业·入门", "cuda": 3328.0, "vramGb": 12.0, "vramPerGpuGb": 12.0, "vramType": "GDDR6", "vramLabel": "12GB/6GB GDDR6", "fp32": null, "date": "2021-06-01", "datePrecision": "year", "dateText": "2021年", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "amp-rtxa1000", "name": "RTX A1000", "arch": "Ampere", "archShort": "AMP", "archYears": "2020—2022", "archTech": "TF32 · MIG", "tier": "workstation", "category": "专业·入门", "cuda": null, "vramGb": 8.0, "vramPerGpuGb": 8.0, "vramType": "GDDR6", "vramLabel": "8GB GDDR6", "fp32": null, "date": "2022-06-01", "datePrecision": "year", "dateText": "2022年", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "amp-rtxa400", "name": "RTX A400", "arch": "Ampere", "archShort": "AMP", "archYears": "2020—2022", "archTech": "TF32 · MIG", "tier": "workstation", "category": "专业·入门", "cuda": null, "vramGb": 4.0, "vramPerGpuGb": 4.0, "vramType": "GDDR6", "vramLabel": "4GB GDDR6", "fp32": null, "date": "2022-06-01", "datePrecision": "year", "dateText": "2022年", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "amp-rtx3090ti", "name": "RTX 3090 Ti", "arch": "Ampere", "archShort": "AMP", "archYears": "2020—2022", "archTech": "TF32 · MIG", "tier": "consumer", "category": "消费级", "cuda": 10752.0, "vramGb": 24.0, "vramPerGpuGb": 24.0, "vramType": "GDDR6X", "vramLabel": "24GB GDDR6X", "fp32": 40.0, "date": "2022-03-01", "datePrecision": "month", "dateText": "2022年3月", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "amp-rtx3090", "name": "RTX 3090", "arch": "Ampere", "archShort": "AMP", "archYears": "2020—2022", "archTech": "TF32 · MIG", "tier": "consumer", "category": "消费级", "cuda": 10496.0, "vramGb": 24.0, "vramPerGpuGb": 24.0, "vramType": "GDDR6X", "vramLabel": "24GB GDDR6X", "fp32": 35.6, "date": "2020-09-01", "datePrecision": "month", "dateText": "2020年9月", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "amp-rtx3080ti", "name": "RTX 3080 Ti", "arch": "Ampere", "archShort": "AMP", "archYears": "2020—2022", "archTech": "TF32 · MIG", "tier": "consumer", "category": "消费级", "cuda": 10240.0, "vramGb": 12.0, "vramPerGpuGb": 12.0, "vramType": "GDDR6X", "vramLabel": "12GB GDDR6X", "fp32": 34.1, "date": "2021-06-01", "datePrecision": "month", "dateText": "2021年6月", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "amp-rtx3080", "name": "RTX 3080", "arch": "Ampere", "archShort": "AMP", "archYears": "2020—2022", "archTech": "TF32 · MIG", "tier": "consumer", "category": "消费级", "cuda": 8704.0, "vramGb": 10.0, "vramPerGpuGb": 10.0, "vramType": "GDDR6X", "vramLabel": "10GB GDDR6X", "fp32": 29.8, "date": "2020-09-01", "datePrecision": "month", "dateText": "2020年9月", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "amp-rtx3070ti", "name": "RTX 3070 Ti", "arch": "Ampere", "archShort": "AMP", "archYears": "2020—2022", "archTech": "TF32 · MIG", "tier": "consumer", "category": "消费级", "cuda": 6144.0, "vramGb": 8.0, "vramPerGpuGb": 8.0, "vramType": "GDDR6X", "vramLabel": "8GB GDDR6X", "fp32": 21.8, "date": "2021-06-01", "datePrecision": "month", "dateText": "2021年6月", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "amp-rtx3070", "name": "RTX 3070", "arch": "Ampere", "archShort": "AMP", "archYears": "2020—2022", "archTech": "TF32 · MIG", "tier": "consumer", "category": "消费级", "cuda": 5888.0, "vramGb": 8.0, "vramPerGpuGb": 8.0, "vramType": "GDDR6", "vramLabel": "8GB GDDR6", "fp32": 20.3, "date": "2020-10-01", "datePrecision": "month", "dateText": "2020年10月", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "amp-rtx3060ti", "name": "RTX 3060 Ti", "arch": "Ampere", "archShort": "AMP", "archYears": "2020—2022", "archTech": "TF32 · MIG", "tier": "consumer", "category": "消费级", "cuda": 4864.0, "vramGb": 8.0, "vramPerGpuGb": 8.0, "vramType": "GDDR6", "vramLabel": "8GB GDDR6", "fp32": 16.2, "date": "2020-12-01", "datePrecision": "month", "dateText": "2020年12月", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "amp-rtx3060", "name": "RTX 3060", "arch": "Ampere", "archShort": "AMP", "archYears": "2020—2022", "archTech": "TF32 · MIG", "tier": "consumer", "category": "消费级", "cuda": 3584.0, "vramGb": 12.0, "vramPerGpuGb": 12.0, "vramType": "GDDR6", "vramLabel": "12GB GDDR6", "fp32": 12.7, "date": "2021-02-01", "datePrecision": "month", "dateText": "2021年2月", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "amp-rtx3050", "name": "RTX 3050", "arch": "Ampere", "archShort": "AMP", "archYears": "2020—2022", "archTech": "TF32 · MIG", "tier": "consumer", "category": "消费级", "cuda": 2560.0, "vramGb": 8.0, "vramPerGpuGb": 8.0, "vramType": "GDDR6", "vramLabel": "8GB GDDR6", "fp32": 9.1, "date": "2022-01-01", "datePrecision": "month", "dateText": "2022年1月", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "tur-teslat4", "name": "Tesla T4", "arch": "Turing", "archShort": "TUR", "archYears": "2018—2020", "archTech": "RT Core · Tensor Core", "tier": "datacenter", "category": "数据中心·推理", "cuda": 2560.0, "vramGb": 16.0, "vramPerGpuGb": 16.0, "vramType": "GDDR6", "vramLabel": "16GB GDDR6", "fp32": 8.1, "date": "2018-09-01", "datePrecision": "month", "dateText": "2018年9月", "note": "70W；FP16张量（稀疏）65 TFLOPS", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "tur-quadrortx8000", "name": "Quadro RTX 8000", "arch": "Turing", "archShort": "TUR", "archYears": "2018—2020", "archTech": "RT Core · Tensor Core", "tier": "workstation", "category": "专业工作站", "cuda": 4608.0, "vramGb": 48.0, "vramPerGpuGb": 48.0, "vramType": "GDDR6", "vramLabel": "48GB GDDR6", "fp32": 16.3, "date": "2018-06-01", "datePrecision": "year", "dateText": "2018年", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "tur-quadrortx6000", "name": "Quadro RTX 6000", "arch": "Turing", "archShort": "TUR", "archYears": "2018—2020", "archTech": "RT Core · Tensor Core", "tier": "workstation", "category": "专业工作站", "cuda": 4608.0, "vramGb": 24.0, "vramPerGpuGb": 24.0, "vramType": "GDDR6", "vramLabel": "24GB GDDR6", "fp32": 16.3, "date": "2018-06-01", "datePrecision": "year", "dateText": "2018年", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "tur-quadrortx5000", "name": "Quadro RTX 5000", "arch": "Turing", "archShort": "TUR", "archYears": "2018—2020", "archTech": "RT Core · Tensor Core", "tier": "workstation", "category": "专业工作站", "cuda": 3072.0, "vramGb": 16.0, "vramPerGpuGb": 16.0, "vramType": "GDDR6", "vramLabel": "16GB GDDR6", "fp32": 11.2, "date": "2018-06-01", "datePrecision": "year", "dateText": "2018年", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "tur-quadrortx4000", "name": "Quadro RTX 4000", "arch": "Turing", "archShort": "TUR", "archYears": "2018—2020", "archTech": "RT Core · Tensor Core", "tier": "workstation", "category": "专业工作站", "cuda": 2304.0, "vramGb": 8.0, "vramPerGpuGb": 8.0, "vramType": "GDDR6", "vramLabel": "8GB GDDR6", "fp32": 7.1, "date": "2018-06-01", "datePrecision": "year", "dateText": "2018年", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "tur-rtx2080ti", "name": "RTX 2080 Ti", "arch": "Turing", "archShort": "TUR", "archYears": "2018—2020", "archTech": "RT Core · Tensor Core", "tier": "consumer", "category": "消费级", "cuda": 4352.0, "vramGb": 11.0, "vramPerGpuGb": 11.0, "vramType": "GDDR6", "vramLabel": "11GB GDDR6", "fp32": 13.4, "date": "2018-09-01", "datePrecision": "month", "dateText": "2018年9月", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "tur-rtx2080super", "name": "RTX 2080 Super", "arch": "Turing", "archShort": "TUR", "archYears": "2018—2020", "archTech": "RT Core · Tensor Core", "tier": "consumer", "category": "消费级", "cuda": 3072.0, "vramGb": 8.0, "vramPerGpuGb": 8.0, "vramType": "GDDR6", "vramLabel": "8GB GDDR6", "fp32": 11.2, "date": "2019-07-01", "datePrecision": "month", "dateText": "2019年7月", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "tur-rtx2080", "name": "RTX 2080", "arch": "Turing", "archShort": "TUR", "archYears": "2018—2020", "archTech": "RT Core · Tensor Core", "tier": "consumer", "category": "消费级", "cuda": 2944.0, "vramGb": 8.0, "vramPerGpuGb": 8.0, "vramType": "GDDR6", "vramLabel": "8GB GDDR6", "fp32": 10.1, "date": "2018-09-01", "datePrecision": "month", "dateText": "2018年9月", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "tur-rtx2070super", "name": "RTX 2070 Super", "arch": "Turing", "archShort": "TUR", "archYears": "2018—2020", "archTech": "RT Core · Tensor Core", "tier": "consumer", "category": "消费级", "cuda": 2560.0, "vramGb": 8.0, "vramPerGpuGb": 8.0, "vramType": "GDDR6", "vramLabel": "8GB GDDR6", "fp32": 9.0, "date": "2019-07-01", "datePrecision": "month", "dateText": "2019年7月", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "tur-rtx2070", "name": "RTX 2070", "arch": "Turing", "archShort": "TUR", "archYears": "2018—2020", "archTech": "RT Core · Tensor Core", "tier": "consumer", "category": "消费级", "cuda": 2304.0, "vramGb": 8.0, "vramPerGpuGb": 8.0, "vramType": "GDDR6", "vramLabel": "8GB GDDR6", "fp32": 7.5, "date": "2018-10-01", "datePrecision": "month", "dateText": "2018年10月", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "tur-rtx2060super", "name": "RTX 2060 Super", "arch": "Turing", "archShort": "TUR", "archYears": "2018—2020", "archTech": "RT Core · Tensor Core", "tier": "consumer", "category": "消费级", "cuda": 2176.0, "vramGb": 8.0, "vramPerGpuGb": 8.0, "vramType": "GDDR6", "vramLabel": "8GB GDDR6", "fp32": 7.2, "date": "2019-07-01", "datePrecision": "month", "dateText": "2019年7月", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "tur-rtx2060", "name": "RTX 2060", "arch": "Turing", "archShort": "TUR", "archYears": "2018—2020", "archTech": "RT Core · Tensor Core", "tier": "consumer", "category": "消费级", "cuda": 1920.0, "vramGb": 6.0, "vramPerGpuGb": 6.0, "vramType": "GDDR6", "vramLabel": "6GB GDDR6", "fp32": 6.5, "date": "2019-01-01", "datePrecision": "month", "dateText": "2019年1月", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "tur-gtx1660ti", "name": "GTX 1660 Ti", "arch": "Turing", "archShort": "TUR", "archYears": "2018—2020", "archTech": "RT Core · Tensor Core", "tier": "consumer", "category": "消费级", "cuda": 1536.0, "vramGb": 6.0, "vramPerGpuGb": 6.0, "vramType": "GDDR6", "vramLabel": "6GB GDDR6", "fp32": 5.4, "date": "2019-02-01", "datePrecision": "month", "dateText": "2019年2月", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "tur-gtx1660super", "name": "GTX 1660 Super", "arch": "Turing", "archShort": "TUR", "archYears": "2018—2020", "archTech": "RT Core · Tensor Core", "tier": "consumer", "category": "消费级", "cuda": 1408.0, "vramGb": 6.0, "vramPerGpuGb": 6.0, "vramType": "GDDR6", "vramLabel": "6GB GDDR6", "fp32": 5.0, "date": "2019-10-01", "datePrecision": "month", "dateText": "2019年10月", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "tur-gtx1660", "name": "GTX 1660", "arch": "Turing", "archShort": "TUR", "archYears": "2018—2020", "archTech": "RT Core · Tensor Core", "tier": "consumer", "category": "消费级", "cuda": 1408.0, "vramGb": 6.0, "vramPerGpuGb": 6.0, "vramType": "GDDR6", "vramLabel": "6GB GDDR6", "fp32": 5.0, "date": "2019-03-01", "datePrecision": "month", "dateText": "2019年3月", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "tur-gtx1650super", "name": "GTX 1650 Super", "arch": "Turing", "archShort": "TUR", "archYears": "2018—2020", "archTech": "RT Core · Tensor Core", "tier": "consumer", "category": "消费级", "cuda": 1280.0, "vramGb": 4.0, "vramPerGpuGb": 4.0, "vramType": "GDDR6", "vramLabel": "4GB GDDR6", "fp32": 4.4, "date": "2019-11-01", "datePrecision": "month", "dateText": "2019年11月", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "tur-gtx1650", "name": "GTX 1650", "arch": "Turing", "archShort": "TUR", "archYears": "2018—2020", "archTech": "RT Core · Tensor Core", "tier": "consumer", "category": "消费级", "cuda": 896.0, "vramGb": 4.0, "vramPerGpuGb": 4.0, "vramType": "GDDR6", "vramLabel": "4GB GDDR6", "fp32": 3.0, "date": "2019-04-01", "datePrecision": "month", "dateText": "2019年4月", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "vol-v100s32gb", "name": "V100S（32GB）", "arch": "Volta", "archShort": "VOL", "archYears": "2017—2019", "archTech": "首代 Tensor Core · HBM2", "tier": "datacenter", "category": "数据中心", "cuda": 5120.0, "vramGb": 32.0, "vramPerGpuGb": 32.0, "vramType": "HBM2", "vramLabel": "32GB HBM2", "fp32": 16.4, "date": "2019-06-01", "datePrecision": "year", "dateText": "2019年", "note": "高频版，2019年推出", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "vol-v100sxm232gb", "name": "V100 SXM2（32GB）", "arch": "Volta", "archShort": "VOL", "archYears": "2017—2019", "archTech": "首代 Tensor Core · HBM2", "tier": "datacenter", "category": "数据中心·HPC/AI", "cuda": 5120.0, "vramGb": 32.0, "vramPerGpuGb": 32.0, "vramType": "HBM2", "vramLabel": "32GB HBM2", "fp32": 15.7, "date": "2018-06-01", "datePrecision": "year", "dateText": "2018年", "note": "FP16张量（稀疏）125 TFLOPS", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "vol-v100sxm216gb", "name": "V100 SXM2（16GB）", "arch": "Volta", "archShort": "VOL", "archYears": "2017—2019", "archTech": "首代 Tensor Core · HBM2", "tier": "datacenter", "category": "数据中心·HPC/AI", "cuda": 5120.0, "vramGb": 16.0, "vramPerGpuGb": 16.0, "vramType": "HBM2", "vramLabel": "16GB HBM2", "fp32": 15.7, "date": "2017-06-01", "datePrecision": "year", "dateText": "2017年", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "vol-v100pcie16gb32gb", "name": "V100 PCIe（16GB/32GB）", "arch": "Volta", "archShort": "VOL", "archYears": "2017—2019", "archTech": "首代 Tensor Core · HBM2", "tier": "datacenter", "category": "数据中心", "cuda": 5120.0, "vramGb": 32.0, "vramPerGpuGb": 32.0, "vramType": "HBM2", "vramLabel": "16GB/32GB HBM2", "fp32": 15.7, "date": "2017-06-01", "datePrecision": "year", "dateText": "2017年", "note": "PCIe版", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "vol-quadrogv100", "name": "Quadro GV100", "arch": "Volta", "archShort": "VOL", "archYears": "2017—2019", "archTech": "首代 Tensor Core · HBM2", "tier": "workstation", "category": "专业工作站", "cuda": 5120.0, "vramGb": 32.0, "vramPerGpuGb": 32.0, "vramType": "HBM2", "vramLabel": "32GB HBM2", "fp32": 15.0, "date": "2018-03-01", "datePrecision": "month", "dateText": "2018年3月", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "vol-titanv", "name": "Titan V", "arch": "Volta", "archShort": "VOL", "archYears": "2017—2019", "archTech": "首代 Tensor Core · HBM2", "tier": "workstation", "category": "半专业/消费", "cuda": 5120.0, "vramGb": 12.0, "vramPerGpuGb": 12.0, "vramType": "HBM2", "vramLabel": "12GB HBM2", "fp32": 15.0, "date": "2017-12-01", "datePrecision": "month", "dateText": "2017年12月", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "vol-titanvceo", "name": "Titan V CEO版", "arch": "Volta", "archShort": "VOL", "archYears": "2017—2019", "archTech": "首代 Tensor Core · HBM2", "tier": "workstation", "category": "半专业/消费", "cuda": 5120.0, "vramGb": 32.0, "vramPerGpuGb": 32.0, "vramType": "HBM2", "vramLabel": "32GB HBM2", "fp32": 15.0, "date": "2018-06-01", "datePrecision": "year", "dateText": "2018年", "note": "限量版", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "pas-teslap100sxm216gb", "name": "Tesla P100 SXM2（16GB）", "arch": "Pascal", "archShort": "PAS", "archYears": "2016—2018", "archTech": "FP16 · HBM2 · 16nm", "tier": "datacenter", "category": "数据中心·HPC/AI", "cuda": 3584.0, "vramGb": 16.0, "vramPerGpuGb": 16.0, "vramType": "HBM2", "vramLabel": "16GB HBM2", "fp32": 9.3, "date": "2016-04-01", "datePrecision": "month", "dateText": "2016年4月", "note": "FP16 18.7 TFLOPS；HPC旗舰", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "pas-teslap100pcie16gb", "name": "Tesla P100 PCIe（16GB）", "arch": "Pascal", "archShort": "PAS", "archYears": "2016—2018", "archTech": "FP16 · HBM2 · 16nm", "tier": "datacenter", "category": "数据中心", "cuda": 3584.0, "vramGb": 16.0, "vramPerGpuGb": 16.0, "vramType": "HBM2", "vramLabel": "16GB HBM2", "fp32": null, "date": "2016-06-01", "datePrecision": "year", "dateText": "2016年", "note": "PCIe版，规格略低", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "pas-teslap40", "name": "Tesla P40", "arch": "Pascal", "archShort": "PAS", "archYears": "2016—2018", "archTech": "FP16 · HBM2 · 16nm", "tier": "datacenter", "category": "数据中心·推理", "cuda": 3840.0, "vramGb": 24.0, "vramPerGpuGb": 24.0, "vramType": "GDDR5X", "vramLabel": "24GB GDDR5X", "fp32": 11.8, "date": "2016-10-01", "datePrecision": "month", "dateText": "2016年10月", "note": "面向FP32推理", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "pas-teslap4", "name": "Tesla P4", "arch": "Pascal", "archShort": "PAS", "archYears": "2016—2018", "archTech": "FP16 · HBM2 · 16nm", "tier": "datacenter", "category": "数据中心·推理", "cuda": 2560.0, "vramGb": 8.0, "vramPerGpuGb": 8.0, "vramType": "GDDR5", "vramLabel": "8GB GDDR5", "fp32": 5.5, "date": "2016-10-01", "datePrecision": "month", "dateText": "2016年10月", "note": "低功耗推理卡", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "pas-quadrop6000", "name": "Quadro P6000", "arch": "Pascal", "archShort": "PAS", "archYears": "2016—2018", "archTech": "FP16 · HBM2 · 16nm", "tier": "workstation", "category": "专业工作站", "cuda": 3840.0, "vramGb": 24.0, "vramPerGpuGb": 24.0, "vramType": "GDDR5X", "vramLabel": "24GB GDDR5X", "fp32": 12.6, "date": "2016-10-01", "datePrecision": "month", "dateText": "2016年10月", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "pas-quadrop5000", "name": "Quadro P5000", "arch": "Pascal", "archShort": "PAS", "archYears": "2016—2018", "archTech": "FP16 · HBM2 · 16nm", "tier": "workstation", "category": "专业工作站", "cuda": 2560.0, "vramGb": 16.0, "vramPerGpuGb": 16.0, "vramType": "GDDR5X", "vramLabel": "16GB GDDR5X", "fp32": 8.9, "date": "2016-10-01", "datePrecision": "month", "dateText": "2016年10月", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "pas-quadrop4000", "name": "Quadro P4000", "arch": "Pascal", "archShort": "PAS", "archYears": "2016—2018", "archTech": "FP16 · HBM2 · 16nm", "tier": "workstation", "category": "专业工作站", "cuda": 1792.0, "vramGb": 8.0, "vramPerGpuGb": 8.0, "vramType": "GDDR5", "vramLabel": "8GB GDDR5", "fp32": 5.3, "date": "2017-06-01", "datePrecision": "year", "dateText": "2017年", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "pas-titanxp", "name": "Titan Xp", "arch": "Pascal", "archShort": "PAS", "archYears": "2016—2018", "archTech": "FP16 · HBM2 · 16nm", "tier": "workstation", "category": "半专业/消费", "cuda": 3840.0, "vramGb": 12.0, "vramPerGpuGb": 12.0, "vramType": "GDDR5X", "vramLabel": "12GB GDDR5X", "fp32": 12.1, "date": "2017-04-01", "datePrecision": "month", "dateText": "2017年4月", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "pas-gtx1080ti", "name": "GTX 1080 Ti", "arch": "Pascal", "archShort": "PAS", "archYears": "2016—2018", "archTech": "FP16 · HBM2 · 16nm", "tier": "consumer", "category": "消费级", "cuda": 3584.0, "vramGb": 11.0, "vramPerGpuGb": 11.0, "vramType": "GDDR5X", "vramLabel": "11GB GDDR5X", "fp32": 11.3, "date": "2017-03-01", "datePrecision": "month", "dateText": "2017年3月", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "pas-titanxpascal", "name": "Titan X（Pascal）", "arch": "Pascal", "archShort": "PAS", "archYears": "2016—2018", "archTech": "FP16 · HBM2 · 16nm", "tier": "consumer", "category": "消费级", "cuda": 3584.0, "vramGb": 12.0, "vramPerGpuGb": 12.0, "vramType": "GDDR5X", "vramLabel": "12GB GDDR5X", "fp32": 10.9, "date": "2016-08-01", "datePrecision": "month", "dateText": "2016年8月", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "pas-gtx1080", "name": "GTX 1080", "arch": "Pascal", "archShort": "PAS", "archYears": "2016—2018", "archTech": "FP16 · HBM2 · 16nm", "tier": "consumer", "category": "消费级", "cuda": 2560.0, "vramGb": 8.0, "vramPerGpuGb": 8.0, "vramType": "GDDR5X", "vramLabel": "8GB GDDR5X", "fp32": 8.9, "date": "2016-05-01", "datePrecision": "month", "dateText": "2016年5月", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "pas-gtx1070ti", "name": "GTX 1070 Ti", "arch": "Pascal", "archShort": "PAS", "archYears": "2016—2018", "archTech": "FP16 · HBM2 · 16nm", "tier": "consumer", "category": "消费级", "cuda": 2432.0, "vramGb": 8.0, "vramPerGpuGb": 8.0, "vramType": "GDDR5", "vramLabel": "8GB GDDR5", "fp32": 8.0, "date": "2017-11-01", "datePrecision": "month", "dateText": "2017年11月", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "pas-gtx1070", "name": "GTX 1070", "arch": "Pascal", "archShort": "PAS", "archYears": "2016—2018", "archTech": "FP16 · HBM2 · 16nm", "tier": "consumer", "category": "消费级", "cuda": 1920.0, "vramGb": 8.0, "vramPerGpuGb": 8.0, "vramType": "GDDR5", "vramLabel": "8GB GDDR5", "fp32": 6.5, "date": "2016-06-01", "datePrecision": "month", "dateText": "2016年6月", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "pas-gtx10606gb", "name": "GTX 1060（6GB）", "arch": "Pascal", "archShort": "PAS", "archYears": "2016—2018", "archTech": "FP16 · HBM2 · 16nm", "tier": "consumer", "category": "消费级", "cuda": 1280.0, "vramGb": 6.0, "vramPerGpuGb": 6.0, "vramType": "GDDR5", "vramLabel": "6GB GDDR5", "fp32": 4.4, "date": "2016-07-01", "datePrecision": "month", "dateText": "2016年7月", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "pas-gtx10603gb", "name": "GTX 1060（3GB）", "arch": "Pascal", "archShort": "PAS", "archYears": "2016—2018", "archTech": "FP16 · HBM2 · 16nm", "tier": "consumer", "category": "消费级", "cuda": 1152.0, "vramGb": 3.0, "vramPerGpuGb": 3.0, "vramType": "GDDR5", "vramLabel": "3GB GDDR5", "fp32": 4.0, "date": "2016-08-01", "datePrecision": "month", "dateText": "2016年8月", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "pas-gtx1050ti", "name": "GTX 1050 Ti", "arch": "Pascal", "archShort": "PAS", "archYears": "2016—2018", "archTech": "FP16 · HBM2 · 16nm", "tier": "consumer", "category": "消费级", "cuda": 768.0, "vramGb": 4.0, "vramPerGpuGb": 4.0, "vramType": "GDDR5", "vramLabel": "4GB GDDR5", "fp32": 2.1, "date": "2016-10-01", "datePrecision": "month", "dateText": "2016年10月", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "pas-gtx1050", "name": "GTX 1050", "arch": "Pascal", "archShort": "PAS", "archYears": "2016—2018", "archTech": "FP16 · HBM2 · 16nm", "tier": "consumer", "category": "消费级", "cuda": 640.0, "vramGb": 2.0, "vramPerGpuGb": 2.0, "vramType": "GDDR5", "vramLabel": "2GB GDDR5", "fp32": 1.9, "date": "2016-10-01", "datePrecision": "month", "dateText": "2016年10月", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "max-teslam40", "name": "Tesla M40", "arch": "Maxwell", "archShort": "MAX", "archYears": "2014—2016", "archTech": "能效跃迁", "tier": "datacenter", "category": "数据中心", "cuda": null, "vramGb": 24.0, "vramPerGpuGb": 24.0, "vramType": "GDDR5", "vramLabel": "24GB GDDR5", "fp32": null, "date": "2015-06-01", "datePrecision": "year", "dateText": "2015", "note": "深度学习/训练", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "max-teslam4", "name": "Tesla M4", "arch": "Maxwell", "archShort": "MAX", "archYears": "2014—2016", "archTech": "能效跃迁", "tier": "datacenter", "category": "数据中心", "cuda": null, "vramGb": 4.0, "vramPerGpuGb": 4.0, "vramType": "GDDR5", "vramLabel": "4GB GDDR5", "fp32": null, "date": "2015-06-01", "datePrecision": "year", "dateText": "2015", "note": "低功耗推理", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "max-teslam10", "name": "Tesla M10", "arch": "Maxwell", "archShort": "MAX", "archYears": "2014—2016", "archTech": "能效跃迁", "tier": "datacenter", "category": "数据中心", "cuda": null, "vramGb": 32.0, "vramPerGpuGb": 8.0, "vramType": "GDDR5", "vramLabel": "4×8GB GDDR5（共32GB）", "fp32": null, "date": "2015-06-01", "datePrecision": "year", "dateText": "2015", "note": "四路GPU，虚拟化", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "max-quadrom6000", "name": "Quadro M6000", "arch": "Maxwell", "archShort": "MAX", "archYears": "2014—2016", "archTech": "能效跃迁", "tier": "workstation", "category": "专业", "cuda": null, "vramGb": 24.0, "vramPerGpuGb": 24.0, "vramType": "GDDR5", "vramLabel": "24GB GDDR5", "fp32": null, "date": "2015-06-01", "datePrecision": "year", "dateText": "2015", "note": "GM200", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "max-quadrom5000", "name": "Quadro M5000", "arch": "Maxwell", "archShort": "MAX", "archYears": "2014—2016", "archTech": "能效跃迁", "tier": "workstation", "category": "专业", "cuda": null, "vramGb": 8.0, "vramPerGpuGb": 8.0, "vramType": "GDDR5", "vramLabel": "8GB GDDR5", "fp32": null, "date": "2015-06-01", "datePrecision": "year", "dateText": "2015", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "max-quadrom4000", "name": "Quadro M4000", "arch": "Maxwell", "archShort": "MAX", "archYears": "2014—2016", "archTech": "能效跃迁", "tier": "workstation", "category": "专业", "cuda": null, "vramGb": 8.0, "vramPerGpuGb": 8.0, "vramType": "GDDR5", "vramLabel": "8GB GDDR5", "fp32": null, "date": "2015-06-01", "datePrecision": "year", "dateText": "2015", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "max-geforcegtxtitanxmaxwell", "name": "GeForce GTX Titan X（Maxwell）", "arch": "Maxwell", "archShort": "MAX", "archYears": "2014—2016", "archTech": "能效跃迁", "tier": "consumer", "category": "消费", "cuda": null, "vramGb": 12.0, "vramPerGpuGb": 12.0, "vramType": "GDDR5", "vramLabel": "12GB GDDR5", "fp32": null, "date": "2015-06-01", "datePrecision": "year", "dateText": "2015", "note": "GM200旗舰", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "max-gtx980ti", "name": "GTX 980 Ti", "arch": "Maxwell", "archShort": "MAX", "archYears": "2014—2016", "archTech": "能效跃迁", "tier": "consumer", "category": "消费", "cuda": null, "vramGb": 6.0, "vramPerGpuGb": 6.0, "vramType": "GDDR5", "vramLabel": "6GB GDDR5", "fp32": null, "date": "2015-06-01", "datePrecision": "year", "dateText": "2015", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "max-gtx980", "name": "GTX 980", "arch": "Maxwell", "archShort": "MAX", "archYears": "2014—2016", "archTech": "能效跃迁", "tier": "consumer", "category": "消费", "cuda": null, "vramGb": 4.0, "vramPerGpuGb": 4.0, "vramType": "GDDR5", "vramLabel": "4GB GDDR5", "fp32": null, "date": "2014-06-01", "datePrecision": "year", "dateText": "2014", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "max-gtx970", "name": "GTX 970", "arch": "Maxwell", "archShort": "MAX", "archYears": "2014—2016", "archTech": "能效跃迁", "tier": "consumer", "category": "消费", "cuda": null, "vramGb": 4.0, "vramPerGpuGb": 4.0, "vramType": "GDDR5", "vramLabel": "4GB GDDR5", "fp32": null, "date": "2014-06-01", "datePrecision": "year", "dateText": "2014", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "max-gtx960", "name": "GTX 960", "arch": "Maxwell", "archShort": "MAX", "archYears": "2014—2016", "archTech": "能效跃迁", "tier": "consumer", "category": "消费", "cuda": null, "vramGb": 4.0, "vramPerGpuGb": 4.0, "vramType": "GDDR5", "vramLabel": "2GB/4GB GDDR5", "fp32": null, "date": "2015-06-01", "datePrecision": "year", "dateText": "2015", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "max-gtx950", "name": "GTX 950", "arch": "Maxwell", "archShort": "MAX", "archYears": "2014—2016", "archTech": "能效跃迁", "tier": "consumer", "category": "消费", "cuda": null, "vramGb": 2.0, "vramPerGpuGb": 2.0, "vramType": "GDDR5", "vramLabel": "2GB GDDR5", "fp32": null, "date": "2015-06-01", "datePrecision": "year", "dateText": "2015", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "max-gtx750ti", "name": "GTX 750 Ti", "arch": "Maxwell", "archShort": "MAX", "archYears": "2014—2016", "archTech": "能效跃迁", "tier": "consumer", "category": "消费", "cuda": null, "vramGb": 2.0, "vramPerGpuGb": 2.0, "vramType": "GDDR5", "vramLabel": "2GB GDDR5", "fp32": null, "date": "2014-06-01", "datePrecision": "year", "dateText": "2014", "note": "GM107，功耗65W", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "kep-teslak80", "name": "Tesla K80", "arch": "Kepler", "archShort": "KEP", "archYears": "2012—2014", "archTech": "GK110 · 动态并行", "tier": "datacenter", "category": "数据中心", "cuda": null, "vramGb": 24.0, "vramPerGpuGb": 12.0, "vramType": "GDDR5", "vramLabel": "2×12GB GDDR5（共24GB）", "fp32": null, "date": "2014-06-01", "datePrecision": "year", "dateText": "2014", "note": "双GPU", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "kep-teslak40", "name": "Tesla K40", "arch": "Kepler", "archShort": "KEP", "archYears": "2012—2014", "archTech": "GK110 · 动态并行", "tier": "datacenter", "category": "数据中心", "cuda": null, "vramGb": 12.0, "vramPerGpuGb": 12.0, "vramType": "GDDR5", "vramLabel": "12GB GDDR5", "fp32": null, "date": "2013-06-01", "datePrecision": "year", "dateText": "2013", "note": "GK110", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "kep-teslak20", "name": "Tesla K20", "arch": "Kepler", "archShort": "KEP", "archYears": "2012—2014", "archTech": "GK110 · 动态并行", "tier": "datacenter", "category": "数据中心", "cuda": null, "vramGb": 5.0, "vramPerGpuGb": 5.0, "vramType": "GDDR5", "vramLabel": "5GB GDDR5", "fp32": null, "date": "2012-06-01", "datePrecision": "year", "dateText": "2012", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "kep-teslak10", "name": "Tesla K10", "arch": "Kepler", "archShort": "KEP", "archYears": "2012—2014", "archTech": "GK110 · 动态并行", "tier": "datacenter", "category": "数据中心", "cuda": null, "vramGb": 8.0, "vramPerGpuGb": 4.0, "vramType": "GDDR5", "vramLabel": "2×4GB GDDR5（共8GB）", "fp32": null, "date": "2012-06-01", "datePrecision": "year", "dateText": "2012", "note": "双GPU", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "kep-quadrok6000", "name": "Quadro K6000", "arch": "Kepler", "archShort": "KEP", "archYears": "2012—2014", "archTech": "GK110 · 动态并行", "tier": "workstation", "category": "专业", "cuda": null, "vramGb": 12.0, "vramPerGpuGb": 12.0, "vramType": "GDDR5", "vramLabel": "12GB GDDR5", "fp32": null, "date": "2013-06-01", "datePrecision": "year", "dateText": "2013", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "kep-quadrok5000", "name": "Quadro K5000", "arch": "Kepler", "archShort": "KEP", "archYears": "2012—2014", "archTech": "GK110 · 动态并行", "tier": "workstation", "category": "专业", "cuda": null, "vramGb": 4.0, "vramPerGpuGb": 4.0, "vramType": "GDDR5", "vramLabel": "4GB GDDR5", "fp32": null, "date": "2012-06-01", "datePrecision": "year", "dateText": "2012", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "kep-quadrok4000", "name": "Quadro K4000", "arch": "Kepler", "archShort": "KEP", "archYears": "2012—2014", "archTech": "GK110 · 动态并行", "tier": "workstation", "category": "专业", "cuda": null, "vramGb": 3.0, "vramPerGpuGb": 3.0, "vramType": "GDDR5", "vramLabel": "3GB GDDR5", "fp32": null, "date": "2013-06-01", "datePrecision": "year", "dateText": "2013", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "kep-geforcegtxtitan", "name": "GeForce GTX Titan", "arch": "Kepler", "archShort": "KEP", "archYears": "2012—2014", "archTech": "GK110 · 动态并行", "tier": "consumer", "category": "消费", "cuda": null, "vramGb": 6.0, "vramPerGpuGb": 6.0, "vramType": "GDDR5", "vramLabel": "6GB GDDR5", "fp32": null, "date": "2013-06-01", "datePrecision": "year", "dateText": "2013", "note": "GK110旗舰", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "kep-gtx780ti", "name": "GTX 780 Ti", "arch": "Kepler", "archShort": "KEP", "archYears": "2012—2014", "archTech": "GK110 · 动态并行", "tier": "consumer", "category": "消费", "cuda": null, "vramGb": 3.0, "vramPerGpuGb": 3.0, "vramType": "GDDR5", "vramLabel": "3GB GDDR5", "fp32": null, "date": "2013-06-01", "datePrecision": "year", "dateText": "2013", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "kep-gtx780", "name": "GTX 780", "arch": "Kepler", "archShort": "KEP", "archYears": "2012—2014", "archTech": "GK110 · 动态并行", "tier": "consumer", "category": "消费", "cuda": null, "vramGb": 3.0, "vramPerGpuGb": 3.0, "vramType": "GDDR5", "vramLabel": "3GB GDDR5", "fp32": null, "date": "2013-06-01", "datePrecision": "year", "dateText": "2013", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "kep-gtx770", "name": "GTX 770", "arch": "Kepler", "archShort": "KEP", "archYears": "2012—2014", "archTech": "GK110 · 动态并行", "tier": "consumer", "category": "消费", "cuda": null, "vramGb": 4.0, "vramPerGpuGb": 4.0, "vramType": "GDDR5", "vramLabel": "2GB/4GB GDDR5", "fp32": null, "date": "2013-06-01", "datePrecision": "year", "dateText": "2013", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "kep-gtx760", "name": "GTX 760", "arch": "Kepler", "archShort": "KEP", "archYears": "2012—2014", "archTech": "GK110 · 动态并行", "tier": "consumer", "category": "消费", "cuda": null, "vramGb": 4.0, "vramPerGpuGb": 4.0, "vramType": "GDDR5", "vramLabel": "2GB/4GB GDDR5", "fp32": null, "date": "2013-06-01", "datePrecision": "year", "dateText": "2013", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "kep-gtx680", "name": "GTX 680", "arch": "Kepler", "archShort": "KEP", "archYears": "2012—2014", "archTech": "GK110 · 动态并行", "tier": "consumer", "category": "消费", "cuda": null, "vramGb": 4.0, "vramPerGpuGb": 4.0, "vramType": "GDDR5", "vramLabel": "2GB/4GB GDDR5", "fp32": null, "date": "2012-06-01", "datePrecision": "year", "dateText": "2012", "note": "GK104", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "kep-gtx670", "name": "GTX 670", "arch": "Kepler", "archShort": "KEP", "archYears": "2012—2014", "archTech": "GK110 · 动态并行", "tier": "consumer", "category": "消费", "cuda": null, "vramGb": 2.0, "vramPerGpuGb": 2.0, "vramType": "GDDR5", "vramLabel": "2GB GDDR5", "fp32": null, "date": "2012-06-01", "datePrecision": "year", "dateText": "2012", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "kep-gtx660ti", "name": "GTX 660 Ti", "arch": "Kepler", "archShort": "KEP", "archYears": "2012—2014", "archTech": "GK110 · 动态并行", "tier": "consumer", "category": "消费", "cuda": null, "vramGb": 3.0, "vramPerGpuGb": 3.0, "vramType": "GDDR5", "vramLabel": "2GB/3GB GDDR5", "fp32": null, "date": "2012-06-01", "datePrecision": "year", "dateText": "2012", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "fer-teslam2090", "name": "Tesla M2090", "arch": "Fermi", "archShort": "FER", "archYears": "2010—2011", "archTech": "首代完整 GPGPU", "tier": "datacenter", "category": "数据中心", "cuda": null, "vramGb": 6.0, "vramPerGpuGb": 6.0, "vramType": "GDDR5", "vramLabel": "6GB GDDR5", "fp32": null, "date": "2011-06-01", "datePrecision": "year", "dateText": "2011", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "fer-teslam2070", "name": "Tesla M2070", "arch": "Fermi", "archShort": "FER", "archYears": "2010—2011", "archTech": "首代完整 GPGPU", "tier": "datacenter", "category": "数据中心", "cuda": null, "vramGb": 6.0, "vramPerGpuGb": 6.0, "vramType": "GDDR5", "vramLabel": "6GB GDDR5", "fp32": null, "date": "2011-06-01", "datePrecision": "year", "dateText": "2011", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "fer-teslam2050", "name": "Tesla M2050", "arch": "Fermi", "archShort": "FER", "archYears": "2010—2011", "archTech": "首代完整 GPGPU", "tier": "datacenter", "category": "数据中心", "cuda": null, "vramGb": 3.0, "vramPerGpuGb": 3.0, "vramType": "GDDR5", "vramLabel": "3GB GDDR5", "fp32": null, "date": "2011-06-01", "datePrecision": "year", "dateText": "2011", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "fer-quadro6000", "name": "Quadro 6000", "arch": "Fermi", "archShort": "FER", "archYears": "2010—2011", "archTech": "首代完整 GPGPU", "tier": "workstation", "category": "专业", "cuda": null, "vramGb": 6.0, "vramPerGpuGb": 6.0, "vramType": "GDDR5", "vramLabel": "6GB GDDR5", "fp32": null, "date": "2010-06-01", "datePrecision": "year", "dateText": "2010", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "fer-quadro5000", "name": "Quadro 5000", "arch": "Fermi", "archShort": "FER", "archYears": "2010—2011", "archTech": "首代完整 GPGPU", "tier": "workstation", "category": "专业", "cuda": null, "vramGb": 2.5, "vramPerGpuGb": 2.5, "vramType": "GDDR5", "vramLabel": "2.5GB GDDR5", "fp32": null, "date": "2010-06-01", "datePrecision": "year", "dateText": "2010", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "fer-quadro4000", "name": "Quadro 4000", "arch": "Fermi", "archShort": "FER", "archYears": "2010—2011", "archTech": "首代完整 GPGPU", "tier": "workstation", "category": "专业", "cuda": null, "vramGb": 2.0, "vramPerGpuGb": 2.0, "vramType": "GDDR5", "vramLabel": "2GB GDDR5", "fp32": null, "date": "2010-06-01", "datePrecision": "year", "dateText": "2010", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "fer-geforcegtx580", "name": "GeForce GTX 580", "arch": "Fermi", "archShort": "FER", "archYears": "2010—2011", "archTech": "首代完整 GPGPU", "tier": "consumer", "category": "消费", "cuda": null, "vramGb": 3.0, "vramPerGpuGb": 3.0, "vramType": "GDDR5", "vramLabel": "1.5GB/3GB GDDR5", "fp32": null, "date": "2010-06-01", "datePrecision": "year", "dateText": "2010", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "fer-gtx570", "name": "GTX 570", "arch": "Fermi", "archShort": "FER", "archYears": "2010—2011", "archTech": "首代完整 GPGPU", "tier": "consumer", "category": "消费", "cuda": null, "vramGb": 1.25, "vramPerGpuGb": 1.25, "vramType": "GDDR5", "vramLabel": "1.25GB GDDR5", "fp32": null, "date": "2010-06-01", "datePrecision": "year", "dateText": "2010", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "fer-gtx560ti", "name": "GTX 560 Ti", "arch": "Fermi", "archShort": "FER", "archYears": "2010—2011", "archTech": "首代完整 GPGPU", "tier": "consumer", "category": "消费", "cuda": null, "vramGb": 2.0, "vramPerGpuGb": 2.0, "vramType": "GDDR5", "vramLabel": "1GB/2GB GDDR5", "fp32": null, "date": "2011-06-01", "datePrecision": "year", "dateText": "2011", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "fer-gtx480", "name": "GTX 480", "arch": "Fermi", "archShort": "FER", "archYears": "2010—2011", "archTech": "首代完整 GPGPU", "tier": "consumer", "category": "消费", "cuda": null, "vramGb": 1.5, "vramPerGpuGb": 1.5, "vramType": "GDDR5", "vramLabel": "1.5GB GDDR5", "fp32": null, "date": "2010-06-01", "datePrecision": "year", "dateText": "2010", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "fer-gtx470", "name": "GTX 470", "arch": "Fermi", "archShort": "FER", "archYears": "2010—2011", "archTech": "首代完整 GPGPU", "tier": "consumer", "category": "消费", "cuda": null, "vramGb": 1.25, "vramPerGpuGb": 1.25, "vramType": "GDDR5", "vramLabel": "1.25GB GDDR5", "fp32": null, "date": "2010-06-01", "datePrecision": "year", "dateText": "2010", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "fer-gtx460", "name": "GTX 460", "arch": "Fermi", "archShort": "FER", "archYears": "2010—2011", "archTech": "首代完整 GPGPU", "tier": "consumer", "category": "消费", "cuda": null, "vramGb": 1.0, "vramPerGpuGb": 1.0, "vramType": "GDDR5", "vramLabel": "1GB GDDR5", "fp32": null, "date": "2010-06-01", "datePrecision": "year", "dateText": "2010", "note": "", "china": false, "brand": "NVIDIA", "source": "docx", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } } ], "extras": [ { "id": "x-7900xtx", "name": "Radeon RX 7900 XTX", "arch": "RDNA 3", "archShort": "RDN", "archYears": "", "archTech": "", "tier": "consumer", "category": "consumer_gpu", "cuda": 6144, "vramGb": 24, "vramType": "GDDR6", "vramLabel": "24GB GDDR6", "fp32": 61.4, "date": "2022-12-13", "datePrecision": "day", "dateText": "2022-12-13", "note": "gpu-server-compare 收录; 非 docx 覆盖型号", "china": false, "brand": "AMD", "source": "gpu-server-compare", "marketId": "7900xtx", "msrpUsd": 999, "priceCny": 7999, "regionAvail": "in_stock", "perfAi": 38, "perfGaming": 66, "tdpW": 355, "bandwidthGbps": 960, "tensorTflops": 122.8, "nvlink": false, "multiMax": 1, "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "x-7900xt", "name": "Radeon RX 7900 XT", "arch": "RDNA 3", "archShort": "RDN", "archYears": "", "archTech": "", "tier": "consumer", "category": "consumer_gpu", "cuda": 5376, "vramGb": 20, "vramType": "GDDR6", "vramLabel": "20GB GDDR6", "fp32": 51.6, "date": "2022-12-13", "datePrecision": "day", "dateText": "2022-12-13", "note": "gpu-server-compare 收录; 非 docx 覆盖型号", "china": false, "brand": "AMD", "source": "gpu-server-compare", "marketId": "7900xt", "msrpUsd": 899, "priceCny": 6499, "regionAvail": "in_stock", "perfAi": 34, "perfGaming": 58, "tdpW": 315, "bandwidthGbps": 800, "tensorTflops": 103.2, "nvlink": false, "multiMax": 1, "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "x-9070xt", "name": "Radeon RX 9070 XT", "arch": "RDNA 4", "archShort": "RDN", "archYears": "", "archTech": "", "tier": "consumer", "category": "consumer_gpu", "cuda": 4096, "vramGb": 16, "vramType": "GDDR6", "vramLabel": "16GB GDDR6", "fp32": 48.7, "date": "2025-03-06", "datePrecision": "day", "dateText": "2025-03-06", "note": "gpu-server-compare 收录; 非 docx 覆盖型号", "china": false, "brand": "AMD", "source": "gpu-server-compare", "marketId": "9070xt", "msrpUsd": 599, "priceCny": 4999, "regionAvail": "in_stock", "perfAi": 44, "perfGaming": 62, "tdpW": 304, "bandwidthGbps": 640, "tensorTflops": 194.8, "nvlink": false, "multiMax": 1, "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "x-9070", "name": "Radeon RX 9070", "arch": "RDNA 4", "archShort": "RDN", "archYears": "", "archTech": "", "tier": "consumer", "category": "consumer_gpu", "cuda": 3584, "vramGb": 16, "vramType": "GDDR6", "vramLabel": "16GB GDDR6", "fp32": 38.3, "date": "2025-03-06", "datePrecision": "day", "dateText": "2025-03-06", "note": "gpu-server-compare 收录; 非 docx 覆盖型号", "china": false, "brand": "AMD", "source": "gpu-server-compare", "marketId": "9070", "msrpUsd": 549, "priceCny": 4599, "regionAvail": "in_stock", "perfAi": 38, "perfGaming": 52, "tdpW": 220, "bandwidthGbps": 640, "tensorTflops": 153.2, "nvlink": false, "multiMax": 1, "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "x-7800xt", "name": "Radeon RX 7800 XT", "arch": "RDNA 3", "archShort": "RDN", "archYears": "", "archTech": "", "tier": "consumer", "category": "consumer_gpu", "cuda": 3840, "vramGb": 16, "vramType": "GDDR6", "vramLabel": "16GB GDDR6", "fp32": 37.3, "date": "2023-09-06", "datePrecision": "day", "dateText": "2023-09-06", "note": "gpu-server-compare 收录; 非 docx 覆盖型号", "china": false, "brand": "AMD", "source": "gpu-server-compare", "marketId": "7800xt", "msrpUsd": 499, "priceCny": 3999, "regionAvail": "in_stock", "perfAi": 28, "perfGaming": 44, "tdpW": 263, "bandwidthGbps": 624, "tensorTflops": 74.6, "nvlink": false, "multiMax": 1, "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "x-b580", "name": "Arc B580", "arch": "Battlemage", "archShort": "BAT", "archYears": "", "archTech": "", "tier": "consumer", "category": "consumer_gpu", "cuda": 2560, "vramGb": 12, "vramType": "GDDR6", "vramLabel": "12GB GDDR6", "fp32": 19.1, "date": "2024-12-13", "datePrecision": "day", "dateText": "2024-12-13", "note": "gpu-server-compare 收录; 非 docx 覆盖型号", "china": false, "brand": "Intel", "source": "gpu-server-compare", "marketId": "b580", "msrpUsd": 249, "priceCny": 1999, "regionAvail": "in_stock", "perfAi": 18, "perfGaming": 26, "tdpW": 190, "bandwidthGbps": 456, "tensorTflops": 38.2, "nvlink": false, "multiMax": 1, "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "x-mi300x", "name": "Instinct MI300X 192GB", "arch": "CDNA 3", "archShort": "CDN", "archYears": "", "archTech": "", "tier": "datacenter", "category": "datacenter_gpu", "cuda": 24320, "vramGb": 192, "vramType": "HBM3", "vramLabel": "192GB HBM3", "fp32": 163.4, "date": "2023-12-06", "datePrecision": "day", "dateText": "2023-12-06", "note": "gpu-server-compare 收录; 非 docx 覆盖型号", "china": false, "brand": "AMD", "source": "gpu-server-compare", "marketId": "mi300x", "msrpUsd": 15000, "priceCny": 125000, "regionAvail": "in_stock", "perfAi": 92, "perfGaming": 0, "tdpW": 750, "bandwidthGbps": 5300, "tensorTflops": 2600, "nvlink": false, "multiMax": 8, "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "x-mi325x", "name": "Instinct MI325X 256GB", "arch": "CDNA 3", "archShort": "CDN", "archYears": "", "archTech": "", "tier": "datacenter", "category": "datacenter_gpu", "cuda": 24320, "vramGb": 256, "vramType": "HBM3e", "vramLabel": "256GB HBM3e", "fp32": 163.4, "date": "2024-10-10", "datePrecision": "day", "dateText": "2024-10-10", "note": "gpu-server-compare 收录; 非 docx 覆盖型号", "china": false, "brand": "AMD", "source": "gpu-server-compare", "marketId": "mi325x", "msrpUsd": 18000, "priceCny": 150000, "regionAvail": "in_stock", "perfAi": 105, "perfGaming": 0, "tdpW": 750, "bandwidthGbps": 6000, "tensorTflops": 2600, "nvlink": false, "multiMax": 8, "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "x-w7900", "name": "Radeon Pro W7900 48GB", "arch": "RDNA 3", "archShort": "RDN", "archYears": "", "archTech": "", "tier": "workstation", "category": "workstation_gpu", "cuda": 6144, "vramGb": 48, "vramType": "GDDR6 ECC", "vramLabel": "48GB GDDR6 ECC", "fp32": 61, "date": "2023-06-13", "datePrecision": "day", "dateText": "2023-06-13", "note": "gpu-server-compare 收录; 非 docx 覆盖型号", "china": false, "brand": "AMD", "source": "gpu-server-compare", "marketId": "w7900", "msrpUsd": 3499, "priceCny": 27999, "regionAvail": "in_stock", "perfAi": 42, "perfGaming": 58, "tdpW": 295, "bandwidthGbps": 960, "tensorTflops": 122, "nvlink": false, "multiMax": 2, "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "x-intel_max1550", "name": "Data Center GPU Max 1550", "arch": "Ponte Vecchio", "archShort": "PON", "archYears": "", "archTech": "", "tier": "datacenter", "category": "datacenter_gpu", "cuda": 8192, "vramGb": 128, "vramType": "HBM2e", "vramLabel": "128GB HBM2e", "fp32": 52, "date": "2023-01-10", "datePrecision": "day", "dateText": "2023-01-10", "note": "gpu-server-compare 收录; 非 docx 覆盖型号", "china": false, "brand": "Intel", "source": "gpu-server-compare", "marketId": "intel_max1550", "msrpUsd": 8000, "priceCny": 65000, "regionAvail": "in_stock", "perfAi": 52, "perfGaming": 0, "tdpW": 600, "bandwidthGbps": 3270, "tensorTflops": 832, "nvlink": false, "multiMax": 8, "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "x-rtx5060ti_8", "name": "GeForce RTX 5060 Ti 8GB", "arch": "Blackwell", "archShort": "BLA", "archYears": "", "archTech": "", "tier": "consumer", "category": "consumer_gpu", "cuda": 4608, "vramGb": 8, "vramType": "GDDR7", "vramLabel": "8GB GDDR7", "fp32": 24.0, "date": "2025-04-16", "datePrecision": "day", "dateText": "2025-04-16", "note": "gpu-server-compare 收录; 非 docx 覆盖型号", "china": false, "brand": "NVIDIA", "source": "gpu-server-compare", "marketId": "rtx5060ti_8", "msrpUsd": 379, "priceCny": 3199, "regionAvail": "in_stock", "perfAi": 34, "perfGaming": 41, "tdpW": 180, "bandwidthGbps": 448, "tensorTflops": 384, "nvlink": false, "multiMax": 1, "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "x-rx9060xt_16", "name": "Radeon RX 9060 XT 16GB", "arch": "RDNA 4", "archShort": "RDN", "archYears": "", "archTech": "", "tier": "consumer", "category": "consumer_gpu", "cuda": 2048, "vramGb": 16, "vramType": "GDDR6", "vramLabel": "16GB GDDR6", "fp32": 25.2, "date": "2025-06-05", "datePrecision": "day", "dateText": "2025-06-05", "note": "gpu-server-compare 收录; 非 docx 覆盖型号", "china": false, "brand": "AMD", "source": "gpu-server-compare", "marketId": "rx9060xt_16", "msrpUsd": 349, "priceCny": 2799, "regionAvail": "in_stock", "perfAi": 32, "perfGaming": 42, "tdpW": 160, "bandwidthGbps": 320, "tensorTflops": 100.8, "nvlink": false, "multiMax": 1, "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "x-rx9060xt_8", "name": "Radeon RX 9060 XT 8GB", "arch": "RDNA 4", "archShort": "RDN", "archYears": "", "archTech": "", "tier": "consumer", "category": "consumer_gpu", "cuda": 2048, "vramGb": 8, "vramType": "GDDR6", "vramLabel": "8GB GDDR6", "fp32": 25.2, "date": "2025-06-05", "datePrecision": "day", "dateText": "2025-06-05", "note": "gpu-server-compare 收录; 非 docx 覆盖型号", "china": false, "brand": "AMD", "source": "gpu-server-compare", "marketId": "rx9060xt_8", "msrpUsd": 279, "priceCny": 2299, "regionAvail": "in_stock", "perfAi": 28, "perfGaming": 40, "tdpW": 150, "bandwidthGbps": 320, "tensorTflops": 100.8, "nvlink": false, "multiMax": 1, "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "x-7900gre", "name": "Radeon RX 7900 GRE", "arch": "RDNA 3", "archShort": "RDN", "archYears": "", "archTech": "", "tier": "consumer", "category": "consumer_gpu", "cuda": 5120, "vramGb": 16, "vramType": "GDDR6", "vramLabel": "16GB GDDR6", "fp32": 46.1, "date": "2024-02-27", "datePrecision": "day", "dateText": "2024-02-27", "note": "gpu-server-compare 收录; 非 docx 覆盖型号", "china": false, "brand": "AMD", "source": "gpu-server-compare", "marketId": "7900gre", "msrpUsd": 549, "priceCny": 3999, "regionAvail": "in_stock", "perfAi": 32, "perfGaming": 52, "tdpW": 260, "bandwidthGbps": 576, "tensorTflops": 92.2, "nvlink": false, "multiMax": 1, "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "x-7700xt", "name": "Radeon RX 7700 XT", "arch": "RDNA 3", "archShort": "RDN", "archYears": "", "archTech": "", "tier": "consumer", "category": "consumer_gpu", "cuda": 3456, "vramGb": 12, "vramType": "GDDR6", "vramLabel": "12GB GDDR6", "fp32": 35.2, "date": "2023-09-06", "datePrecision": "day", "dateText": "2023-09-06", "note": "gpu-server-compare 收录; 非 docx 覆盖型号", "china": false, "brand": "AMD", "source": "gpu-server-compare", "marketId": "7700xt", "msrpUsd": 449, "priceCny": 3199, "regionAvail": "in_stock", "perfAi": 26, "perfGaming": 38, "tdpW": 245, "bandwidthGbps": 432, "tensorTflops": 70.4, "nvlink": false, "multiMax": 1, "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "x-7600xt", "name": "Radeon RX 7600 XT", "arch": "RDNA 3", "archShort": "RDN", "archYears": "", "archTech": "", "tier": "consumer", "category": "consumer_gpu", "cuda": 2048, "vramGb": 16, "vramType": "GDDR6", "vramLabel": "16GB GDDR6", "fp32": 22.6, "date": "2024-01-24", "datePrecision": "day", "dateText": "2024-01-24", "note": "gpu-server-compare 收录; 非 docx 覆盖型号", "china": false, "brand": "AMD", "source": "gpu-server-compare", "marketId": "7600xt", "msrpUsd": 329, "priceCny": 2599, "regionAvail": "in_stock", "perfAi": 18, "perfGaming": 28, "tdpW": 190, "bandwidthGbps": 288, "tensorTflops": 45.2, "nvlink": false, "multiMax": 1, "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "x-b570", "name": "Arc B570", "arch": "Battlemage", "archShort": "BAT", "archYears": "", "archTech": "", "tier": "consumer", "category": "consumer_gpu", "cuda": 2304, "vramGb": 10, "vramType": "GDDR6", "vramLabel": "10GB GDDR6", "fp32": 14.6, "date": "2025-01-16", "datePrecision": "day", "dateText": "2025-01-16", "note": "gpu-server-compare 收录; 非 docx 覆盖型号", "china": false, "brand": "Intel", "source": "gpu-server-compare", "marketId": "b570", "msrpUsd": 219, "priceCny": 1799, "regionAvail": "in_stock", "perfAi": 16, "perfGaming": 22, "tdpW": 150, "bandwidthGbps": 380, "tensorTflops": 29.2, "nvlink": false, "multiMax": 1, "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "x-a770", "name": "Arc A770 16GB", "arch": "Alchemist", "archShort": "ALC", "archYears": "", "archTech": "", "tier": "consumer", "category": "consumer_gpu", "cuda": 4096, "vramGb": 16, "vramType": "GDDR6", "vramLabel": "16GB GDDR6", "fp32": 20.8, "date": "2022-10-12", "datePrecision": "day", "dateText": "2022-10-12", "note": "gpu-server-compare 收录; 非 docx 覆盖型号", "china": false, "brand": "Intel", "source": "gpu-server-compare", "marketId": "a770", "msrpUsd": 349, "priceCny": 2399, "regionAvail": "in_stock", "perfAi": 18, "perfGaming": 28, "tdpW": 225, "bandwidthGbps": 560, "tensorTflops": 41.6, "nvlink": false, "multiMax": 1, "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "x-gh200", "name": "GH200 Grace Hopper 96GB", "arch": "Hopper + Grace", "archShort": "HOP", "archYears": "", "archTech": "", "tier": "datacenter", "category": "datacenter_gpu", "cuda": 16896, "vramGb": 96, "vramType": "HBM3", "vramLabel": "96GB HBM3", "fp32": 66.9, "date": "2023-08-08", "datePrecision": "day", "dateText": "2023-08-08", "note": "gpu-server-compare 收录; 非 docx 覆盖型号", "china": true, "brand": "NVIDIA", "source": "gpu-server-compare", "marketId": "gh200", "msrpUsd": 40000, "priceCny": 340000, "regionAvail": "restricted", "perfAi": 135, "perfGaming": 0, "tdpW": 1000, "bandwidthGbps": 4000, "tensorTflops": 3958, "nvlink": true, "multiMax": 8, "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "x-gb200", "name": "GB200 Grace Blackwell 192GB", "arch": "Blackwell + Grace", "archShort": "BLA", "archYears": "", "archTech": "", "tier": "datacenter", "category": "datacenter_gpu", "cuda": 20864, "vramGb": 192, "vramType": "HBM3e", "vramLabel": "192GB HBM3e", "fp32": 90, "date": "2025-03-18", "datePrecision": "day", "dateText": "2025-03-18", "note": "gpu-server-compare 收录; 非 docx 覆盖型号", "china": true, "brand": "NVIDIA", "source": "gpu-server-compare", "marketId": "gb200", "msrpUsd": 70000, "priceCny": 580000, "regionAvail": "restricted", "perfAi": 220, "perfGaming": 0, "tdpW": 2700, "bandwidthGbps": 8000, "tensorTflops": 10000, "nvlink": true, "multiMax": 72, "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } } ], "servers": [ { "id": "dgx_h100", "name": "NVIDIA DGX H100", "vendor": "NVIDIA", "kind": "server", "gpu": "H100 SXM5 80GB", "gpuId": "h100_sxm", "gpuCount": 8, "powerKw": 10.2, "rackUnits": 6, "segment": "AI Training", "priceUsd": 320000, "priceCny": null, "priceRaw": null, "date": "2022-09-20", "datePrecision": "derived", "dateNote": "随 H100 SXM5 80GB 平台上市", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "dgx_b200", "name": "NVIDIA DGX B200", "vendor": "NVIDIA", "kind": "server", "gpu": "B200 SXM 192GB", "gpuId": "b200_sxm", "gpuCount": 8, "powerKw": 14.3, "rackUnits": 8, "segment": "AI Training", "priceUsd": 480000, "priceCny": null, "priceRaw": null, "date": "2024-10-15", "datePrecision": "derived", "dateNote": "随 B200 SXM 192GB 平台上市", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "smc_8u_h100", "name": "Supermicro SYS-821GE-TNHR 8U (8×H100)", "vendor": "Supermicro", "kind": "server", "gpu": "H100 SXM5 80GB", "gpuId": "h100_sxm", "gpuCount": 8, "powerKw": 10.5, "rackUnits": 8, "segment": "AI Training", "priceUsd": 285000, "priceCny": null, "priceRaw": null, "date": "2022-09-20", "datePrecision": "derived", "dateNote": "随 H100 SXM5 80GB 平台上市", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "dell_xe9680_h100", "name": "Dell PowerEdge XE9680 (8×H100)", "vendor": "Dell", "kind": "server", "gpu": "H100 SXM5 80GB", "gpuId": "h100_sxm", "gpuCount": 8, "powerKw": 10.2, "rackUnits": 6, "segment": "Enterprise AI", "priceUsd": 310000, "priceCny": null, "priceRaw": null, "date": "2022-09-20", "datePrecision": "derived", "dateNote": "随 H100 SXM5 80GB 平台上市", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "smc_mi300x_8u", "name": "Supermicro AS-8125GS-TNHR (8×MI300X)", "vendor": "Supermicro", "kind": "server", "gpu": "Instinct MI300X 192GB", "gpuId": "mi300x", "gpuCount": 8, "powerKw": 12.0, "rackUnits": 8, "segment": "AI Training", "priceUsd": 240000, "priceCny": null, "priceRaw": null, "date": "2023-12-06", "datePrecision": "derived", "dateNote": "随 Instinct MI300X 192GB 平台上市", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "hpe_mi300x", "name": "HPE Cray XD670 (8×MI300X)", "vendor": "HPE", "kind": "server", "gpu": "Instinct MI300X 192GB", "gpuId": "mi300x", "gpuCount": 8, "powerKw": 11.5, "rackUnits": 8, "segment": "HPC / AI", "priceUsd": 255000, "priceCny": null, "priceRaw": null, "date": "2023-12-06", "datePrecision": "derived", "dateNote": "随 Instinct MI300X 192GB 平台上市", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "gigabyte_g593_mi325x", "name": "GIGABYTE G593-ZX1 (8×MI325X)", "vendor": "GIGABYTE", "kind": "server", "gpu": "Instinct MI325X 256GB", "gpuId": "mi325x", "gpuCount": 8, "powerKw": 12.2, "rackUnits": 8, "segment": "AI Training", "priceUsd": 290000, "priceCny": null, "priceRaw": null, "date": "2024-10-10", "datePrecision": "derived", "dateNote": "随 Instinct MI325X 256GB 平台上市", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "smc_l40s_4u", "name": "Supermicro ARS-211M-NR (8×L40S)", "vendor": "Supermicro", "kind": "server", "gpu": "L40S 48GB", "gpuId": "l40s", "gpuCount": 8, "powerKw": 4.8, "rackUnits": 4, "segment": "Inference / Render", "priceUsd": 98000, "priceCny": null, "priceRaw": null, "date": "2023-08-24", "datePrecision": "derived", "dateNote": "随 L40S 48GB 平台上市", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "custom_4090_4u", "name": "Custom 4U (8× RTX 4090) — 非官方", "vendor": "Whitebox", "kind": "server", "gpu": "GeForce RTX 4090", "gpuId": "rtx4090", "gpuCount": 8, "powerKw": 5.2, "rackUnits": 4, "segment": "性价比推理", "priceUsd": 38000, "priceCny": null, "priceRaw": null, "date": "2022-10-12", "datePrecision": "derived", "dateNote": "随 GeForce RTX 4090 平台上市", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "lenovo_sr675_mi300x", "name": "Lenovo ThinkSystem SR675 V3 (4×MI300X)", "vendor": "Lenovo", "kind": "server", "gpu": "Instinct MI300X 192GB", "gpuId": "mi300x", "gpuCount": 4, "powerKw": 6.5, "rackUnits": 3, "segment": "Enterprise AI", "priceUsd": 135000, "priceCny": null, "priceRaw": null, "date": "2023-12-06", "datePrecision": "derived", "dateNote": "随 Instinct MI300X 192GB 平台上市", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "gb200_nvl72", "name": "NVIDIA GB200 NVL72 (72×GB200)", "vendor": "NVIDIA", "kind": "supernode", "gpu": "GB200 Grace Blackwell 192GB", "gpuId": "gb200", "gpuCount": 72, "powerKw": 120.0, "rackUnits": 42, "segment": "AI Training", "priceUsd": 3200000, "priceCny": null, "priceRaw": null, "date": "2025-03-18", "datePrecision": "derived", "dateNote": "随 GB200 Grace Blackwell 192GB 平台上市", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "h100_hgx_server", "name": "NVIDIA HGX H100 8-GPU 整机", "vendor": "多家 (超微/技嘉/戴尔/华硕)", "kind": "server", "gpu": "H100 SXM5 80GB", "gpuId": "h100_sxm", "gpuCount": 8, "powerKw": 9.5, "rackUnits": null, "segment": "AI 训练/推理 · 上一代", "priceUsd": null, "priceCny": 3800000, "priceRaw": "超微H100 380万/台", "date": "2022-09-20", "datePrecision": "derived", "dateNote": "随 H100 SXM5 80GB 平台上市", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "h200_hgx_server", "name": "NVIDIA HGX H200 8-GPU 整机", "vendor": "多家 (超微/戴尔/技嘉/华硕)", "kind": "server", "gpu": "H200 SXM 141GB", "gpuId": "h200_sxm", "gpuCount": 8, "powerKw": 10.5, "rackUnits": null, "segment": "AI 训练/推理 · 主流在售", "priceUsd": null, "priceCny": 6150000, "priceRaw": "超微H200 615万/台", "date": "2024-04-15", "datePrecision": "derived", "dateNote": "随 H200 SXM 141GB 平台上市", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "b300_hgx_server", "name": "NVIDIA HGX B300 8-GPU 整机 (Blackwell Ultra)", "vendor": "多家 (超微/戴尔/技嘉/华硕)", "kind": "server", "gpu": "B300 SXM6 288GB", "gpuId": "b300_sxm", "gpuCount": 8, "powerKw": 13.0, "rackUnits": null, "segment": "AI 训练/推理 · 当期旗舰", "priceUsd": null, "priceCny": 16800000, "priceRaw": "超微B300/4T 1680万/台", "date": "2025-09-01", "datePrecision": "derived", "dateNote": "随 B300 SXM6 (Blackwell Ultra) 平台上市", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "gb300_nvl72", "name": "NVIDIA GB300 NVL72 (超节点整机柜)", "vendor": "Supermicro / NVIDIA 生态", "kind": "supernode", "gpu": "B300 (Blackwell Ultra) 288GB", "gpuId": "b300_sxm", "gpuCount": 72, "powerKw": 135, "rackUnits": 48, "segment": "AI 训练/推理 · 机架级超节点", "priceUsd": null, "priceCny": 127000000, "priceRaw": "超微GB3001.27亿/套", "date": "2025-09-01", "datePrecision": "derived", "dateNote": "随 B300 SXM6 (Blackwell Ultra) 平台上市", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } } ], "models": [ { "id": "deepseek-coder-6.7b-base", "name": "DeepSeek Coder", "org": "deepseek-ai", "orgName": "DeepSeek", "family": "Coder", "tier": "minor", "date": "2023-10-23", "totalParams": "6.7B", "paramsB": 6.7, "activeB": null, "contextLength": "16K", "license": "DeepSeek License", "downloads": 72481, "likes": 711, "breakthrough": null, "innovations": [ "2T tokens 从零训练", "87% 代码语料", "16K 窗口", "FIM 填空补全" ], "summary": "开山之作：从零在 2T tokens（87% 代码）上训练的代码模型家族，6.7B 与 33B 两个尺寸，16K 窗口加填空任务，支持项目级补全。", "headline": null, "paramsSource": "card", "repo": "deepseek-ai/deepseek-coder-6.7b-base", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "deepseek-llm-7b-base", "name": "DeepSeek LLM", "org": "deepseek-ai", "orgName": "DeepSeek", "family": "LLM", "tier": "minor", "date": "2023-11-29", "totalParams": "7B", "paramsB": 7.0, "activeB": null, "contextLength": null, "license": "DeepSeek License", "downloads": 28536, "likes": 353, "breakthrough": null, "innovations": [ "2T 中英 tokens", "稠密 Transformer", "7B / 67B 双尺寸" ], "summary": "初代通用模型，7B 与 67B 双尺寸，2T 中英 tokens 从零训练。此时还是标准稠密 Transformer，架构上尚无独门武器。", "headline": null, "paramsSource": "card", "repo": "deepseek-ai/deepseek-llm-7b-base", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "deepseek-moe-16b-base", "name": "DeepSeekMoE 16B", "org": "deepseek-ai", "orgName": "DeepSeek", "family": "LLM", "tier": "minor", "date": "2024-01-08", "totalParams": "16B", "paramsB": 16.0, "activeB": null, "contextLength": null, "license": "DeepSeek License", "downloads": 11096, "likes": 155, "breakthrough": "细粒度专家 + 共享专家，把「稀疏但有效」这条路走通，成为后续所有 DeepSeek 模型的骨架。", "innovations": [ "DeepSeekMoE", "细粒度专家", "共享专家", "2.8B 激活" ], "summary": "DeepSeekMoE 架构的首次开源落地：细粒度专家切分加共享专家隔离，用约 2.8B 激活参数逼近同规模稠密模型。", "headline": null, "paramsSource": "card", "repo": "deepseek-ai/deepseek-moe-16b-base", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "deepseek-coder-7b-instruct-v1.5", "name": "DeepSeek Coder v1.5", "org": "deepseek-ai", "orgName": "DeepSeek", "family": "Coder", "tier": "minor", "date": "2024-01-25", "totalParams": "7B", "paramsB": 7.0, "activeB": null, "contextLength": "4K", "license": "DeepSeek License", "downloads": 739009, "likes": 160, "breakthrough": null, "innovations": [ "继续预训练 2T", "指令微调 2B", "4K 窗口" ], "summary": "把 DeepSeek-LLM 7B 继续预训练成代码模型，再叠加指令微调。仓库长期是社区最常被拉取的轻量代码模型之一。", "headline": null, "paramsSource": "card", "repo": "deepseek-ai/deepseek-coder-7b-instruct-v1.5", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "deepseek-math-7b-instruct", "name": "DeepSeekMath", "org": "deepseek-ai", "orgName": "DeepSeek", "family": "Math", "tier": "minor", "date": "2024-02-05", "totalParams": "7B", "paramsB": 7.0, "activeB": null, "contextLength": null, "license": "DeepSeek License", "downloads": 5591, "likes": 155, "breakthrough": "GRPO 首次提出：用组内相对优势替代 critic 网络，省掉与策略同规模的价值模型，成为 R1 系列的核心算法。", "innovations": [ "GRPO 算法", "CoT 数学推理", "DeepSeekMath-RL" ], "summary": "专攻数学推理的 7B 模型，要求用思维链并把答案写进 \\boxed{}。模型本身不大，但它背后的算法改写了后来所有推理模型。", "headline": null, "paramsSource": "card", "repo": "deepseek-ai/deepseek-math-7b-instruct", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "deepseek-vl-7b-chat", "name": "DeepSeek-VL", "org": "deepseek-ai", "orgName": "DeepSeek", "family": "VL", "tier": "minor", "date": "2024-03-07", "totalParams": "7B", "paramsB": 7.0, "activeB": null, "contextLength": null, "license": "DeepSeek License", "downloads": 8461, "likes": 272, "breakthrough": null, "innovations": [ "SigLIP-L + SAM-B 混合编码", "1024×1024 输入", "约 400B 图文 token" ], "summary": "首个开源视觉语言模型：SigLIP-L 加 SAM-B 混合视觉编码器，吃 1024×1024 图像，约 400B 图文 token 训练。", "headline": null, "paramsSource": "card", "repo": "deepseek-ai/deepseek-vl-7b-chat", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "deepseek-v2", "name": "DeepSeek-V2", "org": "deepseek-ai", "orgName": "DeepSeek", "family": "LLM", "tier": "flagship", "date": "2024-04-22", "totalParams": "236B", "paramsB": 236.0, "activeB": 21.0, "contextLength": "128K", "license": "DeepSeek License", "downloads": 63813, "likes": 797, "breakthrough": "MLA 多头潜在注意力：把 KV 压进低维潜空间再解压，长上下文显存与带宽成本直接降一个量级，此后成为行业标配。", "innovations": [ "MLA 潜在注意力", "DeepSeekMoE", "8.1T tokens 预训练", "KV cache ↓93.3%" ], "summary": "236B 总参数、21B 激活的 MoE 模型。MLA 与 DeepSeekMoE 同时上线，训练成本降低 42.5%，KV cache 减少 93.3%，最大生成吞吐提升 5.76 倍。", "headline": { "metric": "CMMLU", "score": 84.0 }, "paramsSource": "card", "repo": "deepseek-ai/DeepSeek-V2", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "deepseek-v2-lite", "name": "DeepSeek-V2-Lite", "org": "deepseek-ai", "orgName": "DeepSeek", "family": "LLM", "tier": "major", "date": "2024-05-15", "totalParams": "16B", "paramsB": 16.0, "activeB": 2.4, "contextLength": "32K", "license": "DeepSeek License", "downloads": 342882, "likes": 187, "breakthrough": null, "innovations": [ "16B / 2.4B 激活", "32K 上下文", "单卡 40G 可部署" ], "summary": "把 MLA 与 MoE 缩到 16B 总参数 / 2.4B 激活、32K 上下文，单张 40G 卡即可部署，8×80G 可微调。社区研究 MLA 的入口。", "headline": { "metric": "CMMLU", "score": 64.3 }, "paramsSource": "card", "repo": "deepseek-ai/DeepSeek-V2-Lite", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "deepseek-coder-v2-instruct", "name": "DeepSeek-Coder-V2", "org": "deepseek-ai", "orgName": "DeepSeek", "family": "Coder", "tier": "major", "date": "2024-06-14", "totalParams": "236B", "paramsB": 236.0, "activeB": 21.0, "contextLength": "128K", "license": "DeepSeek License", "downloads": 10491, "likes": 701, "breakthrough": null, "innovations": [ "6T token 继续预训练", "338 种编程语言", "128K 上下文" ], "summary": "在 V2 中间 checkpoint 上继续预训练 6T token，编程语言从 86 种扩到 338 种、上下文从 16K 提到 128K，代码能力对标 GPT4-Turbo。", "headline": null, "paramsSource": "card", "repo": "deepseek-ai/DeepSeek-Coder-V2-Instruct", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "deepseek-v2.5", "name": "DeepSeek-V2.5", "org": "deepseek-ai", "orgName": "DeepSeek", "family": "LLM", "tier": "major", "date": "2024-09-05", "totalParams": "236B", "paramsB": null, "activeB": null, "contextLength": "128K", "license": "DeepSeek License", "downloads": 5831, "likes": 734, "breakthrough": null, "innovations": [ "通用 + 代码合并", "函数调用", "JSON 输出", "FIM 补全" ], "summary": "把 V2-Chat 的通用能力与 Coder-V2 的代码能力合并进一个模型，补齐写作、函数调用与 JSON 输出。V2 与 V3 之间的过渡版本。", "headline": { "metric": "HumanEval python", "score": 89.0 }, "paramsSource": "inferred", "repo": "deepseek-ai/DeepSeek-V2.5", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "deepseek-vl2", "name": "DeepSeek-VL2", "org": "deepseek-ai", "orgName": "DeepSeek", "family": "VL", "tier": "major", "date": "2024-12-13", "totalParams": "27B MoE", "paramsB": null, "activeB": 1.0, "contextLength": null, "license": "DeepSeek License", "downloads": 4874, "likes": 389, "breakthrough": null, "innovations": [ "MoE 视觉语言模型", "动态切图", "视觉定位" ], "summary": "基于 DeepSeekMoE-27B 的 MoE 视觉语言模型，Tiny / Small / VL2 三档分别激活 1.0B / 2.8B / 4.5B，靠动态切图处理高分辨率。", "headline": null, "paramsSource": "inferred", "repo": "deepseek-ai/deepseek-vl2", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "deepseek-v3", "name": "DeepSeek-V3", "org": "deepseek-ai", "orgName": "DeepSeek", "family": "LLM", "tier": "flagship", "date": "2024-12-25", "totalParams": "671B", "paramsB": 671.0, "activeB": 37.0, "contextLength": "128K", "license": "MIT", "downloads": 1067540, "likes": 5887, "breakthrough": "FP8 训练 + 无辅助损失负载均衡 + MTP：把「大模型必须用 BF16 且负载均衡要牺牲性能」这两条经验同时推翻。", "innovations": [ "FP8 混合精度训练", "无辅助损失负载均衡", "MTP 多 token 预测", "14.8T tokens / 2.788M H800h" ], "summary": "671B 总参数 / 37B 激活。首次在超大规模验证 FP8 混合精度训练，引入无辅助损失的负载均衡与多 token 预测，14.8T token 只花 2.788M H800 小时。", "headline": { "metric": "GPQA-Diamond", "score": 59.1 }, "paramsSource": "card", "repo": "deepseek-ai/DeepSeek-V3", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "deepseek-r1", "name": "DeepSeek-R1", "org": "deepseek-ai", "orgName": "DeepSeek", "family": "Reasoning", "tier": "flagship", "date": "2025-01-20", "totalParams": "671B", "paramsB": 671.0, "activeB": 37.0, "contextLength": "128K", "license": "MIT", "downloads": 1241957, "likes": 16194, "breakthrough": "无 SFT 的大规模 RL：自验证、反思与长思维链自发涌现，MIT 协议开放权重，把推理能力变成可复制的开源资产。", "innovations": [ "纯 RL 无 SFT", "冷启动数据", "长思维链", "推理蒸馏" ], "summary": "首个公开验证「纯强化学习即可激发推理能力」的模型系列：R1-Zero 跳过 SFT 直接大规模 RL，R1 加入冷启动数据与两阶段 RL / SFT，并把推理能力蒸馏到 6 个稠密小模型。", "headline": { "metric": "GPQA-Diamond", "score": 71.5 }, "paramsSource": "card", "repo": "deepseek-ai/DeepSeek-R1", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "janus-pro-7b", "name": "Janus-Pro", "org": "deepseek-ai", "orgName": "DeepSeek", "family": "Janus", "tier": "major", "date": "2025-01-26", "totalParams": "7B", "paramsB": null, "activeB": null, "contextLength": null, "license": "MIT (code) / Model License", "downloads": 10897, "likes": 3658, "breakthrough": "解耦视觉编码 + 单塔自回归，让一个模型同时做好图像理解与图像生成。", "innovations": [ "视觉编码解耦", "统一 Transformer", "理解 + 生成同塔" ], "summary": "统一多模态模型：把视觉编码按「理解」与「生成」解耦成两条独立路径，但保留单一自回归 Transformer，缓解同一编码器在两类任务间的冲突。", "headline": null, "paramsSource": "inferred", "repo": "deepseek-ai/Janus-Pro-7B", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "deepseek-v3-0324", "name": "DeepSeek-V3-0324", "org": "deepseek-ai", "orgName": "DeepSeek", "family": "LLM", "tier": "major", "date": "2025-03-24", "totalParams": "671B", "paramsB": null, "activeB": null, "contextLength": "128K", "license": "MIT", "downloads": 934873, "likes": 3169, "breakthrough": "同一套权重、不加参数，只靠后训练就能涨十几分 AIME，证明后训练算力的回报尚未见顶。", "innovations": [ "结构零改动", "后训练涨点", "前端代码生成", "函数调用优化" ], "summary": "结构完全不变的小版本升级，靠后训练把推理、前端页面生成、中文写作与函数调用拉起来：MMLU-Pro、GPQA、AIME、LiveCodeBench 分别 +5.3 / +9.3 / +19.8 / +10.0。", "headline": { "metric": "GPQA-Diamond", "score": 68.4 }, "paramsSource": "inferred", "repo": "deepseek-ai/DeepSeek-V3-0324", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "deepseek-prover-v2-671b", "name": "DeepSeek-Prover-V2", "org": "deepseek-ai", "orgName": "DeepSeek", "family": "Prover", "tier": "major", "date": "2025-04-30", "totalParams": "671B", "paramsB": 671.0, "activeB": null, "contextLength": null, "license": "DeepSeek License", "downloads": 746, "likes": 831, "breakthrough": "神经定理证明首次做出可用的递归分解 + RL 闭环，形式化数学开始进入大模型射程。", "innovations": [ "Lean 4 形式化证明", "子目标递归分解", "二元反馈 RL", "MiniF2F 88.9%" ], "summary": "面向 Lean 4 的形式化定理证明：用 V3 递归拆解子目标，把形式化证明与思维链合成为冷启动数据，再用对错二元反馈做 RL。MiniF2F-test 通过率 88.9%，PutnamBench 解出 658 题中的 49 题。", "headline": { "metric": "MiniF2F-test (pass ratio)", "score": 88.9 }, "paramsSource": "card", "repo": "deepseek-ai/DeepSeek-Prover-V2-671B", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "deepseek-r1-0528", "name": "DeepSeek-R1-0528", "org": "deepseek-ai", "orgName": "DeepSeek", "family": "Reasoning", "tier": "major", "date": "2025-05-28", "totalParams": "671B", "paramsB": null, "activeB": null, "contextLength": "128K", "license": "MIT", "downloads": 225134, "likes": 2460, "breakthrough": "用「想得更久」换分数：思考长度翻倍直接换来 17.5 个 AIME 点，推理强度的可扩展性被明确量化。", "innovations": [ "思考长度 12K→23K", "AIME 2025 87.5%", "幻觉率下降", "支持 system prompt" ], "summary": "R1 的小版本升级：后训练算力加码，AIME 2025 从 70% 升至 87.5%，单题平均思考 token 从 12K 涨到 23K，同时降低幻觉并支持 system prompt。", "headline": { "metric": "GPQA-Diamond", "score": 81.0 }, "paramsSource": "inferred", "repo": "deepseek-ai/DeepSeek-R1-0528", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "deepseek-v3.1", "name": "DeepSeek-V3.1", "org": "deepseek-ai", "orgName": "DeepSeek", "family": "LLM", "tier": "major", "date": "2025-08-21", "totalParams": "671B", "paramsB": 671.0, "activeB": 37.0, "contextLength": "128K", "license": "MIT", "downloads": 233589, "likes": 830, "breakthrough": "一份权重两种模式 + UE8M0 FP8：推理成本与国产芯片适配性同时解决，为后续 Agent 化铺路。", "innovations": [ "混合推理双模式", "UE8M0 FP8 scale", "两阶段长上下文扩展", "Agent 能力增强" ], "summary": "混合推理版本：同一份权重靠切换 chat template 就能在 thinking 与 non-thinking 两档之间切换，工具调用与 agent 能力显著增强，并用 UE8M0 FP8 scale 格式训练权重与激活。", "headline": { "metric": "GPQA-Diamond", "score": 80.1 }, "paramsSource": "card", "repo": "deepseek-ai/DeepSeek-V3.1", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "deepseek-v3.1-terminus", "name": "DeepSeek-V3.1-Terminus", "org": "deepseek-ai", "orgName": "DeepSeek", "family": "LLM", "tier": "major", "date": "2025-09-22", "totalParams": "671B", "paramsB": null, "activeB": null, "contextLength": "128K", "license": "MIT", "downloads": 19156, "likes": 366, "breakthrough": null, "innovations": [ "语言一致性修复", "Code / Search Agent 优化", "搜索工具集更新" ], "summary": "按用户反馈出的修复版：减少中英混输与异常字符，重点优化 Code Agent 与 Search Agent（HLE 15.9→21.7、BrowseComp 30.0→38.5、SimpleQA 93.4→96.8）。", "headline": { "metric": "GPQA-Diamond", "score": 80.7 }, "paramsSource": "inferred", "repo": "deepseek-ai/DeepSeek-V3.1-Terminus", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "deepseek-v3.2-exp", "name": "DeepSeek-V3.2-Exp", "org": "deepseek-ai", "orgName": "DeepSeek", "family": "LLM", "tier": "flagship", "date": "2025-09-29", "totalParams": "671B", "paramsB": null, "activeB": null, "contextLength": "128K", "license": "MIT", "downloads": 60359, "likes": 1000, "breakthrough": "DSA 稀疏注意力：把长文本注意力从「全连接」改成「先选后算」，长上下文训练与推理成本大幅下降而质量几乎不掉。", "innovations": [ "DSA 细粒度稀疏注意力", "Lightning Indexer", "训练配置对齐 V3.1", "长文本效率" ], "summary": "通往下一代架构的实验版：首次引入 DeepSeek Sparse Attention（DSA）做细粒度稀疏注意力。为了公平归因，训练配置刻意与 V3.1-Terminus 对齐，各基准基本持平。", "headline": { "metric": "GPQA-Diamond", "score": 79.9 }, "paramsSource": "inferred", "repo": "deepseek-ai/DeepSeek-V3.2-Exp", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "deepseek-ocr", "name": "DeepSeek-OCR", "org": "deepseek-ai", "orgName": "DeepSeek", "family": "OCR", "tier": "major", "date": "2025-10-17", "totalParams": null, "paramsB": null, "activeB": null, "contextLength": null, "license": "MIT", "downloads": 2394300, "likes": 3351, "breakthrough": "上下文光学压缩：用视觉 token 承载文本，为「不靠拉长 token 序列也能扩上下文」打开一条新路。", "innovations": [ "Contexts Optical Compression", "grounding 转 Markdown", "多档分辨率", "vLLM 官方支持" ], "summary": "提出「上下文光学压缩」：把文本上下文压成视觉 token 再解码，探索视觉通道能装下多少文字。支持 grounding 转 Markdown 与多档分辨率，官方 vLLM 直接支持。", "headline": null, "paramsSource": null, "repo": "deepseek-ai/DeepSeek-OCR", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "deepseek-math-v2", "name": "DeepSeek-Math-V2", "org": "deepseek-ai", "orgName": "DeepSeek", "family": "Math", "tier": "major", "date": "2025-11-27", "totalParams": "671B", "paramsB": null, "activeB": null, "contextLength": "128K", "license": "Apache-2.0", "downloads": 682, "likes": 707, "breakthrough": "生成器与验证器分离并互相放大：把「答案对不对」变成可扩展的监督信号，数学能力随之越过金牌线。", "innovations": [ "自验证数学推理", "LLM 验证器作奖励", "扩展验证算力", "IMO 2025 金牌" ], "summary": "可自验证的数学推理路线：用 LLM 验证器作奖励去训练证明生成器，并靠扩展验证算力自动标注难核验的证明。IMO 2025 与 CMO 2024 达金牌水平，Putnam 2024 拿到 118/120。", "headline": null, "paramsSource": "inferred", "repo": "deepseek-ai/DeepSeek-Math-V2", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "deepseek-v3.2", "name": "DeepSeek-V3.2", "org": "deepseek-ai", "orgName": "DeepSeek", "family": "LLM", "tier": "major", "date": "2025-12-01", "totalParams": "671B", "paramsB": null, "activeB": null, "contextLength": "1M", "license": "MIT", "downloads": 1664551, "likes": 1476, "breakthrough": "大规模 Agentic 数据合成 + 可扩展 RL：把训练信号从「对话」搬到「真实任务环境」，智能体能力成为主线。", "innovations": [ "DSA 稀疏注意力", "可扩展 RL 框架", "Agentic 任务合成", "Speciale 高算力变体" ], "summary": "DSA 稀疏注意力转正，配合可扩展 RL 框架与大规模 Agentic 任务合成管道，主打高效推理与智能体能力；高算力变体 Speciale 称在 IMO / IOI 2025 拿到金牌表现。", "headline": null, "paramsSource": "card", "repo": "deepseek-ai/DeepSeek-V3.2", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "deepseek-ocr-2", "name": "DeepSeek-OCR 2", "org": "deepseek-ai", "orgName": "DeepSeek", "family": "OCR", "tier": "major", "date": "2026-01-27", "totalParams": null, "paramsB": null, "activeB": null, "contextLength": null, "license": "Apache-2.0", "downloads": 942892, "likes": 1089, "breakthrough": "Visual Causal Flow：让视觉编码按因果顺序而非固定栅格推进，向人类阅读方式靠近。", "innovations": [ "Visual Causal Flow", "动态分辨率", "视觉上下文压缩" ], "summary": "延续光学压缩路线，以 Visual Causal Flow 为核心探索更接近人类的视觉编码方式，支持动态分辨率。", "headline": null, "paramsSource": null, "repo": "deepseek-ai/DeepSeek-OCR-2", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "deepseek-v4-flash", "name": "DeepSeek-V4-Flash", "org": "deepseek-ai", "orgName": "DeepSeek", "family": "LLM", "tier": "major", "date": "2026-04-22", "totalParams": "284B", "paramsB": 284.0, "activeB": 13.0, "contextLength": "1M", "license": "MIT", "downloads": 1776693, "likes": 2211, "breakthrough": "CSA + HCA 混合注意力与 mHC：1M 上下文下只需 V3.2 的 27% 单 token 推理 FLOPs 与 10% KV cache。", "innovations": [ "CSA + HCA 混合注意力", "mHC 流形约束超连接", "Muon 优化器", "1M 上下文", "27% FLOPs / 10% KV" ], "summary": "V4 预览版 Flash 型：284B 总参数 / 13B 激活，百万 token 上下文。引入 CSA+HCA 混合注意力、mHC 流形约束超连接与 Muon 优化器。", "headline": { "metric": "GPQA-Diamond", "score": 88.1 }, "paramsSource": "card", "repo": "deepseek-ai/DeepSeek-V4-Flash", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "deepseek-v4-pro", "name": "DeepSeek-V4-Pro", "org": "deepseek-ai", "orgName": "DeepSeek", "family": "LLM", "tier": "flagship", "date": "2026-04-22", "totalParams": "1.6T", "paramsB": 1600.0, "activeB": 49.0, "contextLength": "1M", "license": "MIT", "downloads": 612515, "likes": 5522, "breakthrough": null, "innovations": [ "1.6T 总参数 / 49B 激活", "CSA + HCA", "mHC", "Muon 优化器", "1M 上下文" ], "summary": "V4 预览版 Pro 型：1.6T 总参数 / 49B 激活，百万 token 上下文，与 Flash 共用同一套架构改动，Pro-Max 模式称当时最强开源模型。", "headline": { "metric": "GPQA-Diamond", "score": 90.1 }, "paramsSource": "card", "repo": "deepseek-ai/DeepSeek-V4-Pro", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "deepseek-v4-flash-dspark", "name": "DeepSeek-V4-Flash-DSpark", "org": "deepseek-ai", "orgName": "DeepSeek", "family": "LLM", "tier": "major", "date": "2026-06-27", "totalParams": "284B", "paramsB": 284.0, "activeB": 13.0, "contextLength": "1M", "license": "MIT", "downloads": 697663, "likes": 271, "breakthrough": "DSpark 投机解码模块化：改一个启动参数就能换解码吞吐，推理加速不再绑定权重格式。", "innovations": [ "DSpark 投机解码", "同权重即插即用", "vLLM / SGLang 单参数" ], "summary": "不是新模型，而是给 V4-Flash 同一份权重挂上 DSpark 投机解码模块，vLLM 加 --speculative-config、SGLang 加一个参数即可开启。", "headline": { "metric": "GPQA-Diamond", "score": 88.1 }, "paramsSource": "card", "repo": "deepseek-ai/DeepSeek-V4-Flash-DSpark", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "deepseek-v4-flash-0731", "name": "DeepSeek-V4-Flash-0731", "org": "deepseek-ai", "orgName": "DeepSeek", "family": "LLM", "tier": "major", "date": "2026-07-31", "totalParams": "284B", "paramsB": null, "activeB": null, "contextLength": "1M", "license": "MIT", "downloads": 4393881, "likes": 3931, "breakthrough": "小激活参数反超上一档：13B 激活的 Flash 正式版在终端类 Agent 任务上压过 Pro 预览版。", "innovations": [ "正式版取代预览", "Agentic 能力大幅提升", "DSpark 内置", "reasoning_effort 三档" ], "summary": "Flash 的正式发布版，取代预览版：Agentic 能力大幅提升，并内置 DSpark 投机解码。Terminal Bench 2.1（82.7）与 DeepSWE（54.4）已超过 V4-Pro 预览版，尽管激活参数远小于它。", "headline": { "metric": "Terminal Bench 2.1", "score": 82.7 }, "paramsSource": "card", "repo": "deepseek-ai/DeepSeek-V4-Flash-0731", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "deepseek-v4-pro-0813", "name": "DeepSeek-V4-Pro-0813", "org": "deepseek-ai", "orgName": "DeepSeek", "family": "LLM", "tier": "flagship", "date": "2026-08-13", "totalParams": "1.6T", "paramsB": null, "activeB": null, "contextLength": "1M", "license": "MIT", "downloads": 171782, "likes": 815, "breakthrough": null, "innovations": [ "正式版取代预览", "生产环境强化", "DSpark 内置", "HLE 带工具 60.0" ], "summary": "Pro 的正式发布版，取代预览版，Agentic 能力与生产环境表现显著提升，同样挂载 DSpark。所列基准全面超过 Pro 预览版，Terminal Bench 2.1 达 87.9、HLE 带工具 60.0。", "headline": { "metric": "Terminal Bench 2.1", "score": 87.9 }, "paramsSource": "card", "repo": "deepseek-ai/DeepSeek-V4-Pro-0813", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "deepseek-v4-flash-vision-exp", "name": "DeepSeek-V4-Flash-Vision-Exp", "org": "deepseek-ai", "orgName": "DeepSeek", "family": "VL", "tier": "major", "date": "2026-08-31", "totalParams": "284B", "paramsB": null, "activeB": null, "contextLength": "1M", "license": "MIT", "downloads": 400892, "likes": 862, "breakthrough": null, "innovations": [ "视觉模块 + 继续训练", "DFlash 注意力", "Hyper-Connections", "DSpark 前向" ], "summary": "V4 家族首个实验性多模态模型：在 V4-Flash 架构上加视觉模块继续训练，多模态 Agent 能力明显提升，纯文本 Agent 任务基本持平。", "headline": { "metric": "Terminal Bench 2.1", "score": 83.9 }, "paramsSource": "card", "repo": "deepseek-ai/DeepSeek-V4-Flash-Vision-Exp", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "deepseek-v4.1-flash", "name": "DeepSeek-V4.1-Flash", "org": "deepseek-ai", "orgName": "DeepSeek", "family": "LLM", "tier": "flagship", "date": "2026-09-10", "totalParams": "552B + 196B Engram", "paramsB": 552.0, "activeB": 8.0, "contextLength": "1M", "license": "MIT", "downloads": 6, "likes": 1537, "breakthrough": "CED + CSA2 + FP4 main KV + SWA Bounded Replay：四层同时压缩，全局 KV cache 只剩 890 bytes/token，已是初代模型的 1/437。", "innovations": [ "CED 因果编码器-解码器", "CSA2 跨层共享 KV", "FP4 main KV cache", "SWA Bounded Replay", "Engram 条件记忆", "1M 上下文 / effort 1-100" ], "summary": "多模态 MoE：CED 因果编码器-解码器架构，prefill / decode 分别只激活 8B / 16B。CSA2 加 FP4 main KV cache 把全局 KV 压到约 890 bytes/token，约为 V4-Flash 的四分之一；再叠加 196B Engram 条件记忆与 DSpark。", "headline": { "metric": "GPQA-Diamond", "score": 90.9 }, "paramsSource": "card", "repo": "deepseek-ai/DeepSeek-V4.1-Flash", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "chatglm-6b", "name": "ChatGLM-6B", "org": "zai-org", "orgName": "智谱 GLM", "family": "LLM", "tier": "flagship", "date": "2023-03-13", "totalParams": "62 亿", "paramsB": null, "activeB": null, "contextLength": null, "license": "Apache-2.0（代码）/ Model License（模型权重）", "downloads": 3070, "likes": 2916, "breakthrough": "以 62 亿参数配合 INT4 量化，把中文对话大模型拉到 6GB 显存的消费级显卡上运行。", "innovations": [ "GLM 架构（General Language Model）", "中英双语对话", "监督微调 + 反馈自助 + 人类反馈强化学习(RLHF)", "INT4 量化 6GB 显存消费级部署", "约 1T 中英双语 token 预训练" ], "summary": "智谱开源系列的第一代中英双语对话模型，62 亿参数、GLM 架构，可在消费级显卡本地部署，是整个 ChatGLM/GLM 开源谱系的起点。", "headline": null, "paramsSource": "card", "repo": "zai-org/chatglm-6b", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "chatglm2-6b", "name": "ChatGLM2-6B", "org": "zai-org", "orgName": "智谱 GLM", "family": "LLM", "tier": "flagship", "date": "2023-06-24", "totalParams": "6B", "paramsB": 6.0, "activeB": null, "contextLength": "32K", "license": "Apache-2.0（代码）/ Model License（模型权重）", "downloads": 442087, "likes": 2056, "breakthrough": "上下文长度由初代的 2K 扩展至 32K（对话 8K 训练），同时推理速度提升 42%、INT4 下 6G 显存支持对话长度由 1K 提升到 8K。", "innovations": [ "FlashAttention", "Multi-Query Attention", "上下文 2K→32K", "GLM 混合目标函数", "1.4T 中英双语 token 预训练" ], "summary": "ChatGLM-6B 的第二代版本，借 FlashAttention 与 Multi-Query Attention 把上下文从 2K 提到 32K、推理提速 42%，同尺寸开源模型中竞争力显著增强。", "headline": { "metric": "GSM8K", "score": 571.0 }, "paramsSource": "card", "repo": "zai-org/chatglm2-6b", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "codegeex2-6b", "name": "CodeGeeX2-6B", "org": "zai-org", "orgName": "智谱 GLM", "family": "Coder", "tier": "major", "date": "2023-07-19", "totalParams": "60 亿", "paramsB": null, "activeB": null, "contextLength": "8K", "license": "Apache-2.0（代码）/ Model License（模型权重）", "downloads": 322, "likes": 258, "breakthrough": "仅 60 亿参数即在 HumanEval-X 上超越 150 亿参数的 StarCoder-15B（Python Pass@1 35.9%）。", "innovations": [ "基于 ChatGLM2-6B 架构", "600B 代码数据预训练", "多语言代码生成（插件支持 100+ 语言）", "8192 序列长度", "INT4 量化后仅需 6GB 显存" ], "summary": "CodeGeeX2-6B 是基于 ChatGLM2-6B 架构、经 600B 代码 token 预训练的多语言代码生成模型，Python HumanEval-X Pass@1 达 35.9%，用 6B 参数超过 15B 的 StarCoder-15B。", "headline": { "metric": "HumanEval-X Rust", "score": 321.0 }, "paramsSource": "card", "repo": "zai-org/codegeex2-6b", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "chatglm3-6b", "name": "ChatGLM3-6B", "org": "zai-org", "orgName": "智谱 GLM", "family": "LLM", "tier": "flagship", "date": "2023-10-25", "totalParams": "6B", "paramsB": 6.0, "activeB": null, "contextLength": null, "license": "Apache-2.0（代码）/ Model License（模型权重）", "downloads": 61785, "likes": 1166, "breakthrough": "原生支持工具调用（Function Call）与代码执行，从纯对话模型转向可执行任务的 Agent 形态。", "innovations": [ "全新设计的 Prompt 格式", "原生工具调用（Function Call）", "代码执行（Code Interpreter）", "Agent 任务支持", "10B 以下最强预训练基座（自述，无数字）" ], "summary": "ChatGLM 系列第三代的对话模型，最大变化是采用全新 Prompt 格式并原生支持 Function Call、Code Interpreter 与 Agent 场景，同时开源 Base 与 32K 长文本版本。", "headline": null, "paramsSource": "card", "repo": "zai-org/chatglm3-6b", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "cogvlm-chat-hf", "name": "CogVLM-17B", "org": "zai-org", "orgName": "智谱 GLM", "family": "VL", "tier": "major", "date": "2023-11-16", "totalParams": "17B", "paramsB": 17.0, "activeB": null, "contextLength": null, "license": "Apache-2.0（代码）/ Model License（模型权重）", "downloads": 790, "likes": 199, "breakthrough": "视觉专家模块（Visual Expert）：在预训练语言模型的每一层引入视觉专家，使 VLM 效果超越或持平 PaLI-X 55B（自述）。", "innovations": [ "视觉专家模块（Visual Expert）", "ViT 编码器 + MLP 适配器 + 预训练 LLM 四组件", "跨模态对话与视觉 grounding", "超越或持平 PaLI-X 55B（自述，无数字）" ], "summary": "CogVLM-17B 是开源视觉语言模型，由 10B 视觉参数与 7B 语言参数组成，在 10 个跨模态基准上自称 SOTA，并在 VQAv2/OKVQA/TextVQA/COCO captioning 上排名第二。", "headline": null, "paramsSource": "card", "repo": "zai-org/cogvlm-chat-hf", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "cogagent-chat-hf", "name": "CogAgent-18B", "org": "zai-org", "orgName": "智谱 GLM", "family": "Agent", "tier": "major", "date": "2023-12-15", "totalParams": "18B", "paramsB": 18.0, "activeB": null, "contextLength": null, "license": "Apache-2.0（代码）/ Model License（模型权重）", "downloads": 488, "likes": 68, "breakthrough": "首个把 VLM 用作 GUI Agent：对任意 GUI 截图直接返回任务计划、下一步动作与操作坐标，并支持 1120x1120 超高清输入。", "innovations": [ "高分辨率视觉输入 1120x1120", "GUI Agent：输出计划/下一步动作/坐标", "GUI 问答（网页、PC 应用、移动应用）", "OCR 相关任务增强", "继承 CogVLM 视觉多轮对话与 grounding" ], "summary": "CogAgent-18B 是在 CogVLM 基础上改进的视觉语言模型（11B 视觉 + 7B 语言），主打 GUI Agent、高分辨率图像理解与 OCR 增强，是 cogagent-chat 版本权重。", "headline": null, "paramsSource": "card", "repo": "zai-org/cogagent-chat-hf", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "cogvlm2-llama3-chat-19b", "name": "CogVLM2-LLaMA3-Chat-19B", "org": "zai-org", "orgName": "智谱 GLM", "family": "VL", "tier": "major", "date": "2024-05-16", "totalParams": "19B", "paramsB": 19.0, "activeB": null, "contextLength": "8K", "license": "CogVLM2 LICENSE（other），基于 Meta Llama 3 需同时遵守 LLAMA3_LICENSE", "downloads": 10263, "likes": 220, "breakthrough": "图文文档理解大幅跃升：DocVQA 92.3 超过 GPT-4V(88.4)、OCRBench 756 超过 GPT-4V(656)，且全程不使用外部 OCR 工具。", "innovations": [ "Meta-Llama-3-8B-Instruct 基座", "8K 文本长度", "图像分辨率 1344 * 1344", "中英双语版本（chinese-chat-19B）", "TextVQA/DocVQA 比上一代大幅提升" ], "summary": "CogVLM2 系列基于 Meta-Llama-3-8B-Instruct 构建的 19B 视觉语言模型，支持 8K 文本与 1344×1344 图像输入，DocVQA 92.3、TextVQA 84.2。", "headline": { "metric": "OCRBench", "score": 756.0 }, "paramsSource": "card", "repo": "zai-org/cogvlm2-llama3-chat-19B", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "glm-4-9b-chat", "name": "GLM-4-9B-Chat", "org": "zai-org", "orgName": "智谱 GLM", "family": "LLM", "tier": "flagship", "date": "2024-06-04", "totalParams": "9B", "paramsB": 9.0, "activeB": null, "contextLength": "128K", "license": "GLM-4 License（other）", "downloads": 87488, "likes": 708, "breakthrough": "把 128K 长上下文、26 种多语言与 Function Call 工具调用同时下放到 9B 开源模型（BFCL Overall Acc. 81.00，逼近 gpt-4-turbo 的 81.24）。", "innovations": [ "128K 上下文", "26 种语言多语言支持", "Function Call 自定义工具调用", "网页浏览 + 代码执行", "长文本推理（1M 上下文版本 / 大海捞针实验）" ], "summary": "GLM-4 系列的开源对话模型 GLM-4-9B-Chat，支持 128K 上下文、26 种语言、网页浏览与工具调用，是同代 9B 级开源模型的主力版本。", "headline": { "metric": "XStoryCloze", "score": 90.7 }, "paramsSource": "card", "repo": "zai-org/glm-4-9b-chat", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "glm-4v-9b", "name": "GLM-4V-9B", "org": "zai-org", "orgName": "智谱 GLM", "family": "VL", "tier": "major", "date": "2024-06-04", "totalParams": "9B", "paramsB": 9.0, "activeB": null, "contextLength": "8K", "license": "GLM-4 License（other）", "downloads": 20715, "likes": 268, "breakthrough": "高分辨率文字识别反超闭源旗舰：OCRBench 786 为卡内同表最高（GPT-4o 736、GPT-4v 656），MMBench-EN 81.1 接近 GPT-4o 的 83.4。", "innovations": [ "1120 * 1120 高分辨率输入", "中英双语多轮对话", "文字识别能力（OCRBench 786）", "图表理解（AI2D 81.1）", "8K 上下文" ], "summary": "GLM-4 系列的开源多模态版本，具备 1120×1120 高分辨率中英双语多轮对话能力，卡内表格自述在中英文综合、感知推理、文字识别、图表理解上超越 GPT-4-turbo-2024-04-09、Gemini 1.0 Pro、Qwen-VL-Max 与 Claude 3 Opus。", "headline": { "metric": "MME", "score": 2163.8 }, "paramsSource": "card", "repo": "zai-org/glm-4v-9b", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "cogvideox-2b", "name": "CogVideoX-2B", "org": "zai-org", "orgName": "智谱 GLM", "family": "Video", "tier": "major", "date": "2024-08-05", "totalParams": "2B", "paramsB": 2.0, "activeB": null, "contextLength": null, "license": "Apache-2.0", "downloads": 22018, "likes": 371, "breakthrough": "把文生视频门槛压到消费级：diffusers FP16 最低 4GB、INT8(torchao) 3.6GB 显存即可推理（推理速度 A100 约 90 秒/50 步）。", "innovations": [ "Expert Transformer（论文标题）", "3d_sincos_pos_embed 位置编码", "FP16 训练精度", "diffusers FP16 最低 4GB 显存推理", "720 x 480 / 6 秒 / 8 fps 视频" ], "summary": "CogVideoX 开源文生视频模型的入门版本（2B，FP16 训练），输出 720×480、6 秒、8 fps 视频，主打低成本推理与二次开发。", "headline": null, "paramsSource": "card", "repo": "zai-org/CogVideoX-2b", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "cogvideox-5b", "name": "CogVideoX-5B", "org": "zai-org", "orgName": "智谱 GLM", "family": "Video", "tier": "flagship", "date": "2024-08-17", "totalParams": "5B", "paramsB": 5.0, "activeB": null, "contextLength": null, "license": "CogVideoX LICENSE（other）", "downloads": 15371, "likes": 688, "breakthrough": "用 3d_rope_pos_embed 替换 3d_sincos_pos_embed 以提升生成质量，推理显存最低 5GB（多卡 15GB，A100 约 180 秒/50 步）。", "innovations": [ "3d_rope_pos_embed 位置编码", "BF16 训练精度", "更高视频生成质量与视觉特效（自述）", "diffusers BF16 最低 5GB 显存推理", "支持 INT8 / FP8 量化" ], "summary": "CogVideoX 开源文生视频模型的较大规模版本（5B，BF16 训练），相比 2B 版本提供更高的视频生成质量与视觉效果。", "headline": null, "paramsSource": "card", "repo": "zai-org/CogVideoX-5b", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "glm-4-voice-9b", "name": "GLM-4-Voice-9B", "org": "zai-org", "orgName": "智谱 GLM", "family": "Audio", "tier": "major", "date": "2024-10-24", "totalParams": "9B", "paramsB": 9.0, "activeB": null, "contextLength": null, "license": null, "downloads": 8548, "likes": 120, "breakthrough": "端到端语音对话并可按指令改变情感、语调、语速与方言，无需级联 ASR+TTS。", "innovations": [ "端到端语音理解与生成", "离散化语音 token 的生成与理解", "实时语音对话", "情感/语调/语速/方言可控", "基于 GLM-4-9B 做语音模态预训练与对齐" ], "summary": "GLM-4-Voice 端到端语音模型的 LLM 部分：在 GLM-4-9B 基础上做语音模态预训练与对齐，可直接理解和生成中英文离散化语音并实时对话。", "headline": null, "paramsSource": "card", "repo": "zai-org/glm-4-voice-9b", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "cogvideox1-5-5b", "name": "CogVideoX1.5-5B", "org": "zai-org", "orgName": "智谱 GLM", "family": "Video", "tier": "major", "date": "2024-11-02", "totalParams": "5B", "paramsB": 5.0, "activeB": null, "contextLength": null, "license": "CogVideoX LICENSE（other）", "downloads": 2986, "likes": 77, "breakthrough": "分辨率从 720×480 提升到 1360×768、帧率 8→16 fps，视频时长由 6 秒扩展到 5/10 秒（A100 约 1000 秒/50 步）。", "innovations": [ "分辨率提升至 1360 * 768", "支持 5 秒或 10 秒视频", "16 frames/second", "帧数须为 16N+1（N<=10，默认 81）", "BF16 精度 + INT8/FP8 量化" ], "summary": "CogVideoX 的 1.5 代 5B 文生/图生视频模型（模型卡标注发布日 2024-11-08），分辨率升到 1360×768、帧率 16 fps、支持 5 秒或 10 秒视频。", "headline": null, "paramsSource": "card", "repo": "zai-org/CogVideoX1.5-5B", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "cogview4-6b", "name": "CogView4-6B", "org": "zai-org", "orgName": "智谱 GLM", "family": "Image", "tier": "minor", "date": "2025-03-03", "totalParams": "6B", "paramsB": 6.0, "activeB": null, "contextLength": null, "license": "Apache-2.0", "downloads": 6190, "likes": 257, "breakthrough": "开源文生图首次把中文文字渲染做成主打能力：中文文字 F1 0.6168 / Pick@4 0.3265，DPG-Bench Overall 85.13 领先同表所有对比模型。", "innovations": [ "中英双语文生图（含中文文字渲染）", "DPG-Bench Overall 85.13 超 SDXL/Flux.1-dev/DALL-E 3（卡内表格）", "分辨率 512–2048、可被 32 整除、最大 2^21 px", "基于 GLM-4-9B（base_model）", "BF16/FP32 精度（FP16 不支持，会溢出黑图）" ], "summary": "CogView4-6B 是基于 GLM-4-9B 的中英双语文生图模型（Apache-2.0），卡内 DPG-Bench Overall 85.13 超过 SDXL、Flux.1-dev 与 DALL-E 3，并单独评测了中文文字准确率。", "headline": { "metric": "DPG-Bench Attribute", "score": 91.17 }, "paramsSource": "card", "repo": "zai-org/CogView4-6B", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "glm-4-9b-0414", "name": "GLM-4-9B-0414", "org": "zai-org", "orgName": "智谱 GLM", "family": "LLM", "tier": "major", "date": "2025-04-07", "totalParams": "9B", "paramsB": 9.0, "activeB": null, "contextLength": null, "license": "MIT", "downloads": 20652, "likes": 111, "breakthrough": null, "innovations": [ "GLM-4-0414 系列 9B 稠密版本", "Function Call（JSON 格式工具调用）", "MIT 开源许可", "轻量本地部署定位" ], "summary": "GLM-4-0414 系列的 9B 稠密模型（2025-04 发布，MIT 许可），与 32B 版本共用同一份模型卡文案，卡内评测表只印了 32B 的数据。", "headline": null, "paramsSource": "card", "repo": "zai-org/GLM-4-9B-0414", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "glm-4-32b-0414", "name": "GLM-4-32B-0414", "org": "zai-org", "orgName": "智谱 GLM", "family": "LLM", "tier": "flagship", "date": "2025-04-07", "totalParams": "32B", "paramsB": 32.0, "activeB": null, "contextLength": null, "license": "MIT", "downloads": 4342, "likes": 489, "breakthrough": "32B 稠密模型在 IFEval(87.6)、BFCL-v3(69.6)、TAU-Bench Retail(68.7)、SimpleQA(88.1) 上追平乃至超过 GPT-4o-1120 与 671B 的 DeepSeek-V3-0324，并以 MIT 许可开源。", "innovations": [ "15T 高质量数据预训练（含大量推理型合成数据）", "拒绝采样 + 强化学习后训练", "工程代码与 Artifact 生成", "Function Call / agent 原子能力强化", "SWE-bench 多框架评测（Moatless/Agentless/OpenHands）" ], "summary": "GLM-4-0414 系列的主力 32B 稠密模型（2025-04 发布，MIT 许可），基座 GLM-4-32B-Base-0414 用 15T 数据预训练，卡内自述在多项基准上对标 GPT-4o 与 DeepSeek-V3-0324(671B)。", "headline": { "metric": "SimpleQA", "score": 88.1 }, "paramsSource": "card", "repo": "zai-org/GLM-4-32B-0414", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "glm-z1-32b-0414", "name": "GLM-Z1-32B-0414", "org": "zai-org", "orgName": "智谱 GLM", "family": "Reasoning", "tier": "flagship", "date": "2025-04-08", "totalParams": "32B", "paramsB": 32.0, "activeB": null, "contextLength": "32K", "license": "mit", "downloads": 50636, "likes": 196, "breakthrough": "在 32B 开源规模上做出对标 DeepSeek V3/R1 的深度推理能力，数学与复杂任务较底座显著提升。", "innovations": [ "冷启动 + 扩展强化学习", "数学/代码/逻辑专项训练", "基于成对排序反馈的通用 RL", "YaRN 长上下文外推（factor 4.0）", "强制思考 对话模板" ], "summary": "基于 GLM-4-32B-0414 经冷启动与扩展强化学习打造的 32B 深度思考推理模型。", "headline": null, "paramsSource": "card", "repo": "zai-org/GLM-Z1-32B-0414", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "glm-z1-rumination-32b-0414", "name": "GLM-Z1-Rumination-32B-0414", "org": "zai-org", "orgName": "智谱 GLM", "family": "Reasoning", "tier": "minor", "date": "2025-04-13", "totalParams": "32B", "paramsB": 32.0, "activeB": null, "contextLength": null, "license": "mit", "downloads": 625, "likes": 119, "breakthrough": "把「长时间深度思考 + 边想边搜索」训成一体能力，研究型写作与复杂检索任务显著提升。", "innovations": [ "Rumination 沉思式长思考", "深度思考中集成搜索工具", "多规则奖励端到端 RL", "研究型写作与复杂检索", "内置 search/click/open/finish 函数调用" ], "summary": "对标 OpenAI Deep Research 的沉思式深度推理模型，在长思考过程中调用搜索工具解决开放复杂问题。", "headline": null, "paramsSource": "card", "repo": "zai-org/GLM-Z1-Rumination-32B-0414", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "glm-4-1v-9b-thinking", "name": "GLM-4.1V-9B-Thinking", "org": "zai-org", "orgName": "智谱 GLM", "family": "VL", "tier": "major", "date": "2025-06-28", "totalParams": "9B", "paramsB": 9.0, "activeB": null, "contextLength": "64K", "license": "mit", "downloads": 177072, "likes": 788, "breakthrough": "首个以「思考范式 + 强化学习」训练的 GLM-V 模型，在 28 项基准中 23 项为 10B 级最佳、18 项超越 Qwen-2.5-VL-72B。", "innovations": [ "Thinking 推理范式 + RL", "RLCS 课程采样强化学习", "任意宽高比 / 最高 4K 图像输入", "64k 上下文", "开源 VLM Reward System" ], "summary": "基于 GLM-4-9B-0414 的 9B 视觉语言推理模型，在 10B 级 VLM 中取得 SOTA。", "headline": null, "paramsSource": "card", "repo": "zai-org/GLM-4.1V-9B-Thinking", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "glm-4-5", "name": "GLM-4.5", "org": "zai-org", "orgName": "智谱 GLM", "family": "LLM", "tier": "flagship", "date": "2025-07-20", "totalParams": "355B", "paramsB": 355.0, "activeB": 32.0, "contextLength": "128K", "license": "mit", "downloads": 80489, "likes": 1411, "breakthrough": "以 355B-A32B 规模在 12 项标准基准上取得 63.2 的综合分，位列所有专有与开源模型第 3。", "innovations": [ "355B-A32B MoE 架构", "混合推理（thinking / non-thinking）", "统一推理、编码与 Agent 能力", "MTP 层 + EAGLE 投机解码", "Base 与 FP8 版同步开源" ], "summary": "面向智能体的 355B-A32B MoE 基础模型，统一推理、编码与智能体能力，支持思考/非思考双模式。", "headline": { "metric": "12 项行业标准基准综合平均分", "score": 63.2 }, "paramsSource": "card", "repo": "zai-org/GLM-4.5", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "glm-4-5-air", "name": "GLM-4.5-Air", "org": "zai-org", "orgName": "智谱 GLM", "family": "LLM", "tier": "major", "date": "2025-07-20", "totalParams": "106B", "paramsB": 106.0, "activeB": 12.0, "contextLength": "128K", "license": "mit", "downloads": 101674, "likes": 635, "breakthrough": "以 106B-A12B 的紧凑设计拿到 59.8 的综合分，接近 355B 旗舰而推理成本大幅降低。", "innovations": [ "106B-A12B 紧凑 MoE", "混合推理（thinking / non-thinking）", "高效轻量部署", "MTP 层 + EAGLE 投机解码", "Base 与 FP8 版同步开源" ], "summary": "GLM-4.5 系列中更紧凑高效的 106B-A12B MoE 版本，兼顾性能与部署效率。", "headline": { "metric": "12 项行业标准基准综合平均分", "score": 59.8 }, "paramsSource": "card", "repo": "zai-org/GLM-4.5-Air", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "glm-4-5v", "name": "GLM-4.5V", "org": "zai-org", "orgName": "智谱 GLM", "family": "VL", "tier": "flagship", "date": "2025-08-10", "totalParams": "106B", "paramsB": 106.0, "activeB": 12.0, "contextLength": null, "license": "mit", "downloads": 45375, "likes": 722, "breakthrough": "在 42 项公开视觉语言基准上取得同规模 SOTA，并把 Grounding 与 GUI Agent 能力做进模型。", "innovations": [ "基于 GLM-4.5-Air 的 MoE 视觉语言模型", "42 项视觉语言基准同规模 SOTA", "Thinking 模式开关", "精确 Grounding（bbox）", "GUI / 桌面助手 Agent" ], "summary": "以 GLM-4.5-Air（106B-12B active）为底座的旗舰视觉语言模型，覆盖图像、视频、文档理解与 GUI 操作。", "headline": null, "paramsSource": "card", "repo": "zai-org/GLM-4.5V", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "glm-4-6", "name": "GLM-4.6", "org": "zai-org", "orgName": "智谱 GLM", "family": "LLM", "tier": "flagship", "date": "2025-09-29", "totalParams": null, "paramsB": null, "activeB": null, "contextLength": "200K", "license": "mit", "downloads": 18342, "likes": 1235, "breakthrough": "上下文从 128K 扩到 200K 以支撑更复杂 Agent 任务，在 8 项公开基准上对 GLM-4.5 全面提分。", "innovations": [ "上下文窗口 128K → 200K", "推理中工具调用（TIR）", "搜索 Agent 专用轨迹格式", "更强的 agentic coding", "写作与角色扮演优化" ], "summary": "在 GLM-4.5 基础上把上下文扩到 200K 并全面强化编码、推理、工具与写作的迭代版本。", "headline": null, "paramsSource": null, "repo": "zai-org/GLM-4.6", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "glyph", "name": "Glyph", "org": "zai-org", "orgName": "智谱 GLM", "family": "VL", "tier": "minor", "date": "2025-10-25", "totalParams": null, "paramsB": null, "activeB": null, "contextLength": null, "license": "mit", "downloads": 305, "likes": 74, "breakthrough": "不改 token 级上下文窗口，而是把长上下文转成视觉压缩问题，在保留语义的同时大幅降低计算与显存成本。", "innovations": [ "视觉-文本压缩扩展上下文", "长文本渲染为图像 + VLM 处理", "基于 GLM-4.1V-9B-Base", "把长上下文建模转为多模态问题", "大幅降低计算与显存开销" ], "summary": "把长文本渲染成图像、交由视觉语言模型处理的上下文扩展框架。", "headline": null, "paramsSource": null, "repo": "zai-org/Glyph", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "glm-4-6v-flash", "name": "GLM-4.6V-Flash", "org": "zai-org", "orgName": "智谱 GLM", "family": "VL", "tier": "flagship", "date": "2025-12-07", "totalParams": "9B", "paramsB": 9.0, "activeB": null, "contextLength": "128K", "license": "mit", "downloads": 78394, "likes": 628, "breakthrough": "首次在 GLM-V 系列引入原生 Function Calling，把「视觉感知」直接接到「可执行动作」。", "innovations": [ "原生多模态 Function Calling", "图文交错内容生成", "128k 多文档/长文档理解", "前端截图还原 HTML/CSS", "GLM-V 系列 9B 轻量版" ], "summary": "GLM-4.6V 系列的 9B 轻量多模态模型，面向本地部署与低延迟场景。", "headline": null, "paramsSource": "card", "repo": "zai-org/GLM-4.6V-Flash", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "autoglm-phone-9b", "name": "AutoGLM-Phone-9B", "org": "zai-org", "orgName": "智谱 GLM", "family": "Agent", "tier": "major", "date": "2025-12-08", "totalParams": "9B", "paramsB": 9.0, "activeB": null, "contextLength": null, "license": "mit", "downloads": 2563, "likes": 444, "breakthrough": "把视觉语言模型直接接进真实手机 GUI：自然语言下达任务即可自动解析意图、规划并执行操作。", "innovations": [ "手机 GUI Agent 模型", "多模态屏幕理解 + 智能规划", "ADB 远程控制设备", "敏感操作确认 / 人工兜底", "MobileRL 在线智能体强化学习" ], "summary": "AutoGLM 手机智能助手框架中的开源模型，理解手机屏幕并自动完成跨 App 操作任务。", "headline": null, "paramsSource": "card", "repo": "zai-org/AutoGLM-Phone-9B", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "glm-asr-nano-2512", "name": "GLM-ASR-Nano-2512", "org": "zai-org", "orgName": "智谱 GLM", "family": "Audio", "tier": "major", "date": "2025-12-09", "totalParams": "1.5B", "paramsB": 1.5, "activeB": null, "contextLength": null, "license": "mit", "downloads": 56071, "likes": 389, "breakthrough": "以 1.5B 规模取得同类开源模型最低平均错误率 4.10，并在多基准上超过 OpenAI Whisper V3。", "innovations": [ "1.5B 语音识别模型", "粤语等方言优化", "低音量/耳语语音鲁棒性", "中文基准领先（Wenet Meeting、Aishell-1）", "支持 transformers / vLLM / SGLang" ], "summary": "1.5B 的开源语音识别模型，重点优化方言与低音量场景。", "headline": { "metric": "Average error rate", "score": 4.1 }, "paramsSource": "card", "repo": "zai-org/GLM-ASR-Nano-2512", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "glm-4-7", "name": "GLM-4.7", "org": "zai-org", "orgName": "智谱 GLM", "family": "LLM", "tier": "flagship", "date": "2025-12-22", "totalParams": null, "paramsB": null, "activeB": null, "contextLength": null, "license": "mit", "downloads": 75379, "likes": 2054, "breakthrough": "SWE-bench Verified 73.8%（+5.8）、Terminal Bench 2.0 41.0%（+16.5）、HLE w/ Tools 42.8%（+12.4）较 GLM-4.6 大幅提升，并新增 Preserved Thinking。", "innovations": [ "Interleaved Thinking", "Preserved Thinking（跨轮保留思考块）", "Turn-level Thinking（逐轮思考开关）", "多语言 agentic coding", "工具使用（τ²-Bench / BrowseComp）" ], "summary": "面向编码与 Agent 的旗舰模型，强化 UI/Vibe Coding、工具使用与复杂推理能力。", "headline": { "metric": "HMMT Feb. 2025", "score": 97.1 }, "paramsSource": null, "repo": "zai-org/GLM-4.7", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "glm-image", "name": "GLM-Image", "org": "zai-org", "orgName": "智谱 GLM", "family": "Image", "tier": "major", "date": "2026-01-08", "totalParams": null, "paramsB": null, "activeB": null, "contextLength": null, "license": "mit", "downloads": 10117, "likes": 1107, "breakthrough": "CVTG-2K 词准确率 0.9116 / NED 0.9557、LongText-Bench ZH 0.979，文本渲染能力显著领先同表开源与闭源模型。", "innovations": [ "自回归 + 扩散混合解码器架构", "9B 自回归生成器（GLM-4-9B-0414 初始化）", "7B 单流 DiT 扩散解码器", "Glyph Encoder 文本渲染模块", "解耦式 GRPO 后训练" ], "summary": "混合自回归 + 扩散解码器的图像生成模型，在文本渲染与知识密集型生成上优势明显。", "headline": null, "paramsSource": null, "repo": "zai-org/GLM-Image", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "glm-4-7-flash", "name": "GLM-4.7-Flash", "org": "zai-org", "orgName": "智谱 GLM", "family": "LLM", "tier": "flagship", "date": "2026-01-19", "totalParams": "30B", "paramsB": 30.0, "activeB": 3.0, "contextLength": null, "license": "mit", "downloads": 1882455, "likes": 1838, "breakthrough": "以 30B-A3B 在 SWE-bench Verified 59.2、τ²-Bench 79.5 上大幅领先同量级 Qwen3-30B-A3B-Thinking-2507 与 GPT-OSS-20B。", "innovations": [ "30B-A3B MoE 架构", "30B 级最强（模型卡自述）", "Preserved Thinking", "MTP 投机解码", "轻量部署（vLLM / SGLang）" ], "summary": "30B-A3B 的 MoE 轻量模型，主打 30B 级别性能与效率平衡。", "headline": { "metric": "AIME 25", "score": 91.6 }, "paramsSource": "card", "repo": "zai-org/GLM-4.7-Flash", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "glm-ocr", "name": "GLM-OCR", "org": "zai-org", "orgName": "智谱 GLM", "family": "OCR", "tier": "flagship", "date": "2026-01-30", "totalParams": "0.9B", "paramsB": 0.9, "activeB": null, "contextLength": null, "license": "mit", "downloads": 1889384, "likes": 2032, "breakthrough": "以 0.9B 规模在 OmniDocBench V1.5 取得 94.62 分排名第一，并可在 vLLM/SGLang/Ollama 上高并发部署。", "innovations": [ "Multi-Token Prediction (MTP) loss", "全任务稳定强化学习", "CogViT 视觉编码器", "GLM-0.5B 语言解码器 + 轻量跨模态连接器", "PP-DocLayout-V3 布局分析 + 并行识别两阶段流水线" ], "summary": "0.9B 参数的多模态 OCR 模型，面向复杂文档解析与结构化信息抽取。", "headline": { "metric": "OmniDocBench V1.5", "score": 94.62 }, "paramsSource": "card", "repo": "zai-org/GLM-OCR", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "glm-5", "name": "GLM-5", "org": "zai-org", "orgName": "智谱 GLM", "family": "LLM", "tier": "flagship", "date": "2026-02-11", "totalParams": "744B", "paramsB": 744.0, "activeB": 40.0, "contextLength": "200K", "license": "mit", "downloads": 58754, "likes": 2119, "breakthrough": "从 GLM-4.5 的 355B-A32B 扩到 744B-A40B、预训练数据从 23T 增到 28.5T，并在推理/编码/Agent 上取得全球开源最佳，逼近前沿闭源模型。", "innovations": [ "744B-A40B MoE", "DeepSeek Sparse Attention (DSA)", "slime 异步 RL 基础设施", "预训练数据 23T → 28.5T tokens", "MTP 投机解码（num_speculative_tokens 3）" ], "summary": "面向复杂系统工程与长程 Agent 任务的 744B-A40B 开源旗舰模型。", "headline": { "metric": "HMMT Nov. 2025", "score": 96.9 }, "paramsSource": "card", "repo": "zai-org/GLM-5", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "glm-5-1", "name": "GLM-5.1", "org": "zai-org", "orgName": "智谱 GLM", "family": "LLM", "tier": "major", "date": "2026-04-03", "totalParams": null, "paramsB": null, "activeB": null, "contextLength": null, "license": "mit", "downloads": 82309, "likes": 1841, "breakthrough": "在长程任务上可维持数百轮优化、数千次工具调用持续提分（SWE-Bench Pro 58.4 领先 GLM-5、Terminal-Bench 2.0 63.5 vs GLM-5 的 56.2）。", "innovations": [ "长程 agentic 工程能力", "SWE-Bench Pro 领先", "数百轮迭代持续优化", "数千次工具调用不退化", "SGLang/vLLM/xLLM/KTransformers 多框架支持" ], "summary": "面向 agentic 工程的下一代旗舰模型，重点提升长程任务中的持续优化能力。", "headline": { "metric": "AIME 2026", "score": 95.3 }, "paramsSource": null, "repo": "zai-org/GLM-5.1", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "scail-2", "name": "SCAIL-2", "org": "zai-org", "orgName": "智谱 GLM", "family": "Video", "tier": "minor", "date": "2026-06-09", "totalParams": null, "paramsB": null, "activeB": null, "contextLength": null, "license": "mit", "downloads": 6445, "likes": 333, "breakthrough": "去掉骨架/掩码中间表示实现端到端驱动，并涌现出跨身份换人、动物驱动、多角色与 SAM3D-Body 零样本控制等能力。", "innovations": [ "端到端受控角色动画（无中间姿态表征）", "Unified Motion Transfer Interface + 专用掩码通道与 RoPE", "6 万 motion pairs 合成数据训练", "反向驱动训练配方", "跨身份换人 / 动物驱动 / SAM3D-Body 零样本" ], "summary": "端到端受控角色动画模型，用驱动视频直接驱动参考角色，不依赖骨架或掩码等中间表征。", "headline": null, "paramsSource": null, "repo": "zai-org/SCAIL-2", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "glm-5-2", "name": "GLM-5.2", "org": "zai-org", "orgName": "智谱 GLM", "family": "LLM", "tier": "flagship", "date": "2026-06-16", "totalParams": null, "paramsB": null, "activeB": null, "contextLength": "1M", "license": "mit", "downloads": 932248, "likes": 5084, "breakthrough": "solid 1M 上下文 + IndexShare 使 1M 长度下每 token FLOPs 降 2.9×，HLE 40.5、AIME 2026 99.2 大幅超越 GLM-5.1。", "innovations": [ "solid 1M token 上下文", "IndexShare（每 4 层稀疏注意力复用同一 indexer）", "每 token FLOPs 降低 2.9×", "MTP 层改进，接受长度最高 +20%", "思考强度（effort）分级" ], "summary": "首个能在 1M token 上下文上稳定支撑长程任务的旗舰模型，并提出 IndexShare 稀疏注意力。", "headline": { "metric": "AIME 2026", "score": 99.2 }, "paramsSource": null, "repo": "zai-org/GLM-5.2", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "glm-5-3", "name": "GLM-5.3", "org": "zai-org", "orgName": "智谱 GLM", "family": "LLM", "tier": "flagship", "date": "2026-08-25", "totalParams": null, "paramsB": null, "activeB": null, "contextLength": "1M", "license": "other (glm-5.3)", "downloads": 635504, "likes": 1811, "breakthrough": "编码能力较 GLM-5.2 提升 50%（自研 Z.ai Code Bench），CyberGym 84.5 居 SOTA，ExploitGym 2h/6h 达 105/130（GLM-5.2 为 29/39）。", "innovations": [ "与 GLM-5.2 同基座，增益全部来自后训练", "Z.ai Code Bench 较 GLM-5.2 提升 50%", "Terminal Bench 3.0 / Agents' Last Exam 开源 SOTA", "涌现网络攻防能力（CyberGym / ExploitGym）", "reasoning_effort 三档思考预算（low/high/max）" ], "summary": "与 GLM-5.2 共用基座、所有提升来自后训练的编码与长程任务强化版。", "headline": { "metric": "GDPval-AA v2", "score": 1769.0 }, "paramsSource": null, "repo": "zai-org/GLM-5.3", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "glm-5-3-flash", "name": "GLM-5.3-Flash", "org": "zai-org", "orgName": "智谱 GLM", "family": "LLM", "tier": "flagship", "date": "2026-08-25", "totalParams": "320B", "paramsB": 320.0, "activeB": 18.0, "contextLength": null, "license": "mit", "downloads": 1333574, "likes": 2269, "breakthrough": "首次在 GLM 系列引入稀疏+线性混合注意力与 mHC 超连接，320B-18B 以十分之一价格超过 GLM-5.2 并逼近 Claude Opus 4.8 的编码/Agent 水平。", "innovations": [ "GLM-5 系列首个原生多模态模型", "稀疏 + 线性注意力混合架构", "Manifold-Constrained Hyper-Connections (mHC)", "30T token 多模态预训练语料", "价格约为 GLM-5.2 的十分之一" ], "summary": "GLM-5 系列首个原生多模态模型，320B-A18B，主打低成本高性能。", "headline": null, "paramsSource": "card", "repo": "zai-org/GLM-5.3-Flash", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "moonlight-16b-a3b-instruct", "name": "Moonlight-16B-A3B-Instruct", "org": "moonshotai", "orgName": "月之暗面 Kimi", "family": "LLM", "tier": "flagship", "date": "2025-02-22", "totalParams": "16B", "paramsB": 16.0, "activeB": 3.0, "contextLength": "8K", "license": "mit", "downloads": 58542, "likes": 204, "breakthrough": "首次证明 Muon 优化器可用于大规模 LLM 训练：约 2x 样本效率、约 52% 训练 FLOPs 达到 AdamW 同等性能", "innovations": [ "Muon 优化器", "权重衰减(Weight Decay)", "一致 RMS 更新", "ZeRO-1 分布式 Muon", "3B/16B MoE" ], "summary": "用 Muon 优化器训练的 3B/16B MoE 模型，5.7T token 预训练，验证了 Muon 在大规模训练上的可扩展性。", "headline": { "metric": "CMath", "score": 81.1 }, "paramsSource": "card", "repo": "moonshotai/Moonlight-16B-A3B-Instruct", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "kimi-vl-a3b-instruct", "name": "Kimi-VL-A3B-Instruct", "org": "moonshotai", "orgName": "月之暗面 Kimi", "family": "VL", "tier": "flagship", "date": "2025-04-09", "totalParams": "16B", "paramsB": 16.0, "activeB": 3.0, "contextLength": "128K", "license": "mit", "downloads": 162717, "likes": 283, "breakthrough": "以 2.8B+0.4B 激活参数在 ScreenSpot-V2(92.8)、OSWorld(8.22) 等 GUI Agent 任务上超过 GPT-4o", "innovations": [ "MoE 视觉语言模型", "MoonViT 原生分辨率视觉编码器", "MLP 投影器", "128K 长上下文", "GUI/OS Agent 能力" ], "summary": "激活仅 2.8B 参数的高效 MoE VLM（16B 总参/3B 激活），128K 上下文，具备长视频/长文档与 OS Agent 能力。", "headline": { "metric": "OCRBench", "score": 867.0 }, "paramsSource": "card", "repo": "moonshotai/Kimi-VL-A3B-Instruct", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "kimi-vl-a3b-thinking", "name": "Kimi-VL-A3B-Thinking", "org": "moonshotai", "orgName": "月之暗面 Kimi", "family": "VL", "tier": "minor", "date": "2025-04-09", "totalParams": "16B", "paramsB": 16.0, "activeB": 3.0, "contextLength": "128K", "license": "mit", "downloads": 16242, "likes": 454, "breakthrough": "以 2.8B 激活参数在多模态数学推理上匹配 30B/70B 级开源 VLM（MathVision 36.8）", "innovations": [ "长思维链(CoT) SFT", "强化学习(RL)", "MoE 视觉语言模型", "2.8B 激活 LLM" ], "summary": "Kimi-VL 的长思考变体，经长 CoT SFT + RL 训练，在 2.8B 激活参数下取得强多模态推理能力。", "headline": { "metric": "MathVista", "score": 71.3 }, "paramsSource": "card", "repo": "moonshotai/Kimi-VL-A3B-Thinking", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "moonvit-so-400m", "name": "MoonViT-SO-400M", "org": "moonshotai", "orgName": "月之暗面 Kimi", "family": "VL", "tier": "major", "date": "2025-04-10", "totalParams": null, "paramsB": null, "activeB": null, "contextLength": null, "license": "mit", "downloads": 2784, "likes": 95, "breakthrough": null, "innovations": [ "原生分辨率视觉编码器", "由 SigLIP-SO-400M 初始化并继续预训练", "与 Kimi-VL 解耦独立发布" ], "summary": "从 Kimi-VL-A3B-Instruct 中拆出的原生分辨率视觉编码器，初始化自 SigLIP-SO-400M；模型卡未印参数量，故参数字段为 null。", "headline": null, "paramsSource": null, "repo": "moonshotai/MoonViT-SO-400M", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "kimi-audio-7b-instruct", "name": "Kimi-Audio-7B-Instruct", "org": "moonshotai", "orgName": "月之暗面 Kimi", "family": "Audio", "tier": "major", "date": "2025-04-25", "totalParams": null, "paramsB": null, "activeB": null, "contextLength": null, "license": "mit", "downloads": 12991, "likes": 417, "breakthrough": null, "innovations": [ "混合音频输入（连续声学+离散语义 token）", "并行文本/音频输出头", "flow matching 流式检测器", "1300万小时音频预训练" ], "summary": "统一的音频基础模型，覆盖 ASR/AQA/AAC/SER/SEC 与端到端语音对话；模型卡未印参数量（仅模型名含 7B），故参数字段为 null。", "headline": null, "paramsSource": null, "repo": "moonshotai/Kimi-Audio-7B-Instruct", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "kimi-dev-72b", "name": "Kimi-Dev-72B", "org": "moonshotai", "orgName": "月之暗面 Kimi", "family": "Coder", "tier": "major", "date": "2025-06-16", "totalParams": null, "paramsB": null, "activeB": null, "contextLength": null, "license": "mit", "downloads": 3628, "likes": 393, "breakthrough": "SWE-bench Verified 达 60.4%，刷新开源模型 SOTA", "innovations": [ "大规模强化学习", "Docker 中真实仓库打补丁", "全测试套件通过才给奖励" ], "summary": "面向软件工程问题解决的编码 LLM（base_model 为 Qwen/Qwen2.5-72B）；模型卡未印参数量，故参数字段为 null。", "headline": { "metric": "SWE-bench Verified", "score": 60.4 }, "paramsSource": null, "repo": "moonshotai/Kimi-Dev-72B", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "kimi-vl-a3b-thinking-2506", "name": "Kimi-VL-A3B-Thinking-2506", "org": "moonshotai", "orgName": "月之暗面 Kimi", "family": "VL", "tier": "major", "date": "2025-06-21", "totalParams": "16B", "paramsB": 16.0, "activeB": 3.0, "contextLength": "128K", "license": "mit", "downloads": 29685, "likes": 383, "breakthrough": "多模态推理大幅提升并降低 20% 思维长度：MathVision 36.8→56.9，单图支持 3.2M 像素（4x）", "innovations": [ "更少 think token(+准确率)", "单图 3.2M 像素高分辨率", "视频推理增强", "MoE 视觉语言模型" ], "summary": "Kimi-VL-A3B-Thinking 的升级版：base_model 指向 Kimi-VL-A3B-Instruct（同代 16B 总参/3B 激活），上下文长度按卡内 VLLM 示例 max_model_len=131072(=128K) 记为 128K；本卡未印参数量表，参数与上下文档为同代推断。", "headline": { "metric": "OCRBench", "score": 869.0 }, "paramsSource": "inferred", "repo": "moonshotai/Kimi-VL-A3B-Thinking-2506", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "kimi-k2-base", "name": "Kimi-K2-Base", "org": "moonshotai", "orgName": "月之暗面 Kimi", "family": "LLM", "tier": "major", "date": "2025-07-03", "totalParams": "1T", "paramsB": 1000.0, "activeB": 32.0, "contextLength": "128K", "license": "modified-mit", "downloads": 12671, "likes": 306, "breakthrough": "MuonClip 优化器把 Muon 首次扩展到 1T 参数 MoE 规模，实现 15.5T token 零训练不稳定", "innovations": [ "MuonClip 优化器", "MoE 1T 总参/32B 激活", "MLA 注意力", "384 专家选 8", "15.5T token 预训练零不稳定" ], "summary": "Kimi K2 的基座模型：1T 总参/32B 激活的 MoE，用 Muon 优化器在 15.5T token 上预训练且训练全程无失稳。", "headline": { "metric": "C-Eval", "score": 92.5 }, "paramsSource": "card", "repo": "moonshotai/Kimi-K2-Base", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "kimi-k2-instruct", "name": "Kimi-K2-Instruct", "org": "moonshotai", "orgName": "月之暗面 Kimi", "family": "LLM", "tier": "flagship", "date": "2025-07-11", "totalParams": "1T", "paramsB": 1000.0, "activeB": 32.0, "contextLength": "128K", "license": "modified-mit", "downloads": 184577, "likes": 2377, "breakthrough": "开源 1T 参数级 agentic 模型：SWE-bench Verified 65.8%（agentic 单次）、Tau2/AceBench 等工具使用基准领先", "innovations": [ "MuonClip 优化器", "MoE 1T 总参/32B 激活", "MLA 注意力", "Agentic 工具调用", "block-fp8 权重" ], "summary": "K2 的后训练版（非长思考的 reflex 级模型），1T 总参/32B 激活 MoE，专为工具使用与自主问题求解优化。", "headline": { "metric": "MATH-500", "score": 97.4 }, "paramsSource": "card", "repo": "moonshotai/Kimi-K2-Instruct", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "kimi-k2-instruct-0905", "name": "Kimi-K2-Instruct-0905", "org": "moonshotai", "orgName": "月之暗面 Kimi", "family": "LLM", "tier": "minor", "date": "2025-09-03", "totalParams": "1T", "paramsB": 1000.0, "activeB": 32.0, "contextLength": "256K", "license": "modified-mit", "downloads": 37788, "likes": 786, "breakthrough": "上下文从 128K 提升到 256K，SWE-bench Verified 由 65.8 提升至 69.2", "innovations": [ "256K 长上下文(由128K扩展)", "Agentic coding 增强", "前端代码生成", "MoE 1T/32B 激活", "MLA" ], "summary": "Kimi K2 系列最新版指令模型，1T 总参/32B 激活，上下文由 128K 扩至 256K，主打 agentic coding。", "headline": { "metric": "SWE-Bench verified", "score": 69.2 }, "paramsSource": "card", "repo": "moonshotai/Kimi-K2-Instruct-0905", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "kimi-linear-48b-a3b-instruct", "name": "Kimi-Linear-48B-A3B-Instruct", "org": "moonshotai", "orgName": "月之暗面 Kimi", "family": "LLM", "tier": "flagship", "date": "2025-10-30", "totalParams": "48B", "paramsB": 48.0, "activeB": 3.0, "contextLength": "1M", "license": "mit", "downloads": 173293, "likes": 595, "breakthrough": "KDA 混合线性注意力：1M 上下文解码吞吐最高 6x 提升、KV cache 需求降低最多 75%，RULER(128k) 84.3", "innovations": [ "Kimi Delta Attention (KDA)", "3:1 KDA-to-global MLA 混合", "Gated DeltaNet 改进的门控", "开源 KDA kernel (FLA)", "KV cache 降低 75%" ], "summary": "混合线性注意力模型（KDA:MLA=3:1），5.7T token 训练，长上下文下相比全注意力最多 6x 解码吞吐提升。", "headline": { "metric": "RULER", "score": 84.3 }, "paramsSource": "card", "repo": "moonshotai/Kimi-Linear-48B-A3B-Instruct", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "kimi-k2-thinking", "name": "Kimi-K2-Thinking", "org": "moonshotai", "orgName": "月之暗面 Kimi", "family": "LLM", "tier": "flagship", "date": "2025-11-04", "totalParams": "1T", "paramsB": 1000.0, "activeB": 32.0, "contextLength": "256K", "license": "modified-mit", "downloads": 51157, "likes": 1714, "breakthrough": "QAT 原生 INT4 量化实现约 2x 生成加速且性能无损，并在 HLE/BrowseComp 等基准刷新 SOTA", "innovations": [ "原生 INT4 量化(QAT)", "交错思维链+工具调用", "200-300 步长程工具编排", "MoE 1T/32B 激活", "MLA" ], "summary": "K2 的开源思考模型，边推理边动态调用工具，256K 上下文，采用 QAT 原生 INT4 量化。", "headline": { "metric": "AIME25", "score": 94.5 }, "paramsSource": "card", "repo": "moonshotai/Kimi-K2-Thinking", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "kimi-k2-5", "name": "Kimi K2.5", "org": "moonshotai", "orgName": "月之暗面 Kimi", "family": "LLM", "tier": "flagship", "date": "2026-01-01", "totalParams": "1T", "paramsB": 1000.0, "activeB": 32.0, "contextLength": "256K", "license": "modified-mit", "downloads": 380587, "likes": 2869, "breakthrough": "原生多模态 + Agent Swarm：BrowseComp 由 60.6 提升到 74.9（上下文管理）、78.4（Agent Swarm）", "innovations": [ "原生多模态（约15T 图文 token 继续预训练）", "Agent Swarm 自协调并行子代理", "MoonViT 400M 视觉编码器", "原生 INT4 量化", "256K 上下文" ], "summary": "在 Kimi-K2-Base 上以约 15 万亿图文混合 token 继续预训练得到的原生多模态 agentic 模型，含 Instant/Thinking 双模式与 Agent Swarm。", "headline": { "metric": "AIME 2025", "score": 96.1 }, "paramsSource": "card", "repo": "moonshotai/Kimi-K2.5", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "kimi-k2-6", "name": "Kimi K2.6", "org": "moonshotai", "orgName": "月之暗面 Kimi", "family": "LLM", "tier": "flagship", "date": "2026-04-14", "totalParams": "1T", "paramsB": 1000.0, "activeB": 32.0, "contextLength": "256K", "license": "modified-mit", "downloads": 535907, "likes": 1605, "breakthrough": "Agent Swarm 横向扩展到 300 个子代理、4000 步协同，BrowseComp(Agent Swarm) 86.3、SWE-Bench Verified 80.2", "innovations": [ "Agent Swarm 扩展至 300 子代理/4000 步", "长程编码（Rust/Go/Python）", "编码驱动设计", "原生 INT4 量化", "原生多模态" ], "summary": "K2.5 的后继原生多模态 agentic 模型，强化长程编码、编码驱动设计与 300 子代理的 Agent Swarm 编排。", "headline": { "metric": "V*", "score": 96.9 }, "paramsSource": "card", "repo": "moonshotai/Kimi-K2.6", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "kimi-k2-7-code", "name": "Kimi K2.7 Code", "org": "moonshotai", "orgName": "月之暗面 Kimi", "family": "Coder", "tier": "major", "date": "2026-06-11", "totalParams": "1T", "paramsB": 1000.0, "activeB": 32.0, "contextLength": "256K", "license": "modified-mit", "downloads": 182708, "likes": 1387, "breakthrough": "相比 K2.6 思考 token 用量减少约 30%：Kimi Code Bench v2 50.9→62.0、MCP Mark Verified 72.8→81.1", "innovations": [ "基于 K2.6 的编码专用 agentic 模型", "原生 INT4 量化", "强制 preserve_thinking", "thinking token 减少约 30%", "MoonViT 视觉编码器" ], "summary": "基于 Kimi K2.6 打造的编码专用 agentic 模型，强化长程工程任务端到端完成能力，并提升 token 效率。", "headline": { "metric": "MCP Mark Verified", "score": 81.1 }, "paramsSource": "card", "repo": "moonshotai/Kimi-K2.7-Code", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "kimi-k3", "name": "Kimi K3", "org": "moonshotai", "orgName": "月之暗面 Kimi", "family": "LLM", "tier": "flagship", "date": "2026-06-13", "totalParams": "2.8T", "paramsB": 2800.0, "activeB": 104.0, "contextLength": "1M", "license": "kimi-k3", "downloads": 2277273, "likes": 11299, "breakthrough": "以 KDA+Attention Residuals+Stable LatentMoE（896 专家激活 16）把 MoE 稀疏度推高，整体扩展效率较 Kimi K2 提升约 2.5x", "innovations": [ "Kimi Delta Attention (KDA)", "Attention Residuals (AttnRes)", "Stable LatentMoE（896 专家激活 16）", "MXFP4权重/MXFP8激活 QAT", "百万 token 上下文" ], "summary": "全球首个开源的 3T 级模型：2.8T 总参/104B 激活，基于 KDA + AttnRes + Stable LatentMoE，原生多模态，1M token 上下文（模型卡写作 1048576）。", "headline": { "metric": "GDPval-AA v2", "score": 1686.0 }, "paramsSource": "card", "repo": "moonshotai/Kimi-K3", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } } ], "breakthroughs": [ { "id": "deepseekmoe", "label": "DeepSeekMoE", "name": "细粒度专家 + 共享专家", "date": "2024-01-08", "lane": "架构", "org": "deepseek-ai", "orgName": "DeepSeek", "releaseId": "deepseek-moe-16b-base", "impact": "同样算力下专家组合空间更大、通用知识又不被稀释，成为之后所有 DeepSeek 模型的骨架。", "magnitude": "16B 总参数 / 2.8B 激活", "mechanism": "把 FFN 切成更多更小的专家，再额外留出始终激活的共享专家隔离通用知识，路由器只需在细碎专家间做选择。", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "grpo", "label": "GRPO", "name": "组相对策略优化", "date": "2024-02-05", "lane": "训练", "org": "deepseek-ai", "orgName": "DeepSeek", "releaseId": "deepseek-math-7b-instruct", "impact": "RL 训练显存与工程复杂度大降，后来成为 R1 系列在超大规模上做推理 RL 的核心算法。", "magnitude": "去掉 critic 网络，RL 成本约减半", "mechanism": "对同一问题采样一组回答，用组内相对得分当优势，直接丢掉与策略同规模的价值网络。", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "mla", "label": "MLA", "name": "多头潜在注意力", "date": "2024-04-22", "lane": "架构", "org": "deepseek-ai", "orgName": "DeepSeek", "releaseId": "deepseek-v2", "impact": "长上下文推理的显存与带宽开销直接降一个量级，此后被各大开源模型广泛采用。", "magnitude": "KV cache ↓93.3%，吞吐 ×5.76", "mechanism": "把 K/V 投影到低维潜向量缓存，推理时再解压成多头，缓存的是压缩后的潜变量而不是完整 KV。", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "aux-loss-free", "label": "Aux-loss-free", "name": "无辅助损失负载均衡", "date": "2024-12-25", "lane": "训练", "org": "deepseek-ai", "orgName": "DeepSeek", "releaseId": "deepseek-v3", "impact": "负载均衡不再以牺牲模型质量为代价，MoE 训练少了一个长期存在的两难。", "magnitude": "均衡与性能不再互斥", "mechanism": "用可学习的偏置项调节专家路由概率来均衡负载，而不是在损失里加一项惩罚。", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "fp8-training", "label": "FP8", "name": "超大规模 FP8 混合精度训练", "date": "2024-12-25", "lane": "训练", "org": "deepseek-ai", "orgName": "DeepSeek", "releaseId": "deepseek-v3", "impact": "首个在 671B 规模跑通的 FP8 训练，把单位 token 的训练成本压下来，也让后来 UE8M0 路线成为可能。", "magnitude": "14.8T token / 2.788M H800 小时", "mechanism": "矩阵乘法用 FP8、累积与关键算子保持高精度，配合细粒度量化与在线 scale 更新稳定训练。", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "mtp", "label": "MTP", "name": "多 token 预测", "date": "2024-12-25", "lane": "推理", "org": "deepseek-ai", "orgName": "DeepSeek", "releaseId": "deepseek-v3", "impact": "既提升训练信号密度，又白送一个推理加速器，是「训练设计即推理优化」的典型例子。", "magnitude": "V3 论文报告解码吞吐约 ×1.8", "mechanism": "训练时让模型一次预测多个未来 token，推理时把这些头改造成投机解码的草稿器。", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "pure-rl", "label": "RL w/o SFT", "name": "纯强化学习激发推理", "date": "2025-01-20", "lane": "后训练", "org": "deepseek-ai", "orgName": "DeepSeek", "releaseId": "deepseek-r1", "impact": "自验证、反思、长思维链自发涌现，首次证明推理能力可以「训出来」而不是「教出来」。", "magnitude": "推理能力由 RL 自发涌现", "mechanism": "直接在基座模型上做大规模 RL，不先做 SFT，只靠可验证答案的对错作奖励。", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "reasoning-distill", "label": "Distill", "name": "推理能力蒸馏到稠密模型", "date": "2025-01-20", "lane": "后训练", "org": "deepseek-ai", "orgName": "DeepSeek", "releaseId": "deepseek-r1", "impact": "推理能力不再是大模型的专属，32B 稠密模型即可对标当时的闭源推理小模型。", "magnitude": "32B 蒸馏模型 AIME 2024 pass@1 72.6%", "mechanism": "用大模型生成的 80 万条推理数据去微调 Qwen / Llama 稠密模型，让小模型继承长链推理。", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "hybrid-thinking", "label": "Hybrid Think", "name": "混合推理双模式", "date": "2025-08-21", "lane": "推理", "org": "deepseek-ai", "orgName": "DeepSeek", "releaseId": "deepseek-v3.1", "impact": "成本与质量变成同一条连续带上的选择，同时为国产芯片的 microscaling 格式预留了接口。", "magnitude": "一份权重两档模式", "mechanism": "同一份权重通过 chat template 切换 thinking / non-thinking，配合 UE8M0 FP8 scale 训练权重与激活。", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "dsa", "label": "DSA", "name": "细粒度稀疏注意力", "date": "2025-09-29", "lane": "架构", "org": "deepseek-ai", "orgName": "DeepSeek", "releaseId": "deepseek-v3.2-exp", "impact": "长文本注意力从全连接改成「先选后算」，训练与推理的长上下文成本大幅下降而质量基本不掉。", "magnitude": "长文本效率跃升，基准与 V3.1 持平", "mechanism": "用 lightning indexer 先给每个 query 选出最相关的 top-k token，注意力只在被选中的部分计算。", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "optical-compression", "label": "OCR Context", "name": "上下文光学压缩", "date": "2025-10-17", "lane": "多模态", "org": "deepseek-ai", "orgName": "DeepSeek", "releaseId": "deepseek-ocr", "impact": "为「上下文不是越长越好、而是换一种表示」打开新路，OCR 从工具变成上下文压缩器。", "magnitude": "等量文本用更少的视觉 token", "mechanism": "把文字渲染成图再编码成视觉 token，用远少于文本 token 的数量承载同样的信息。", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "self-verifiable-math", "label": "Verifier", "name": "可自验证数学推理", "date": "2025-11-27", "lane": "后训练", "org": "deepseek-ai", "orgName": "DeepSeek", "releaseId": "deepseek-math-v2", "impact": "把「答案对不对」变成可扩展的监督信号，数学能力随之越过竞赛金牌线。", "magnitude": "IMO 2025 金牌，Putnam 2024 118/120", "mechanism": "证明生成器与 LLM 验证器分开训练，验证器当奖励模型，并靠扩展验证算力自动标注难核验的证明。", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "agentic-synthesis", "label": "Agentic RL", "name": "大规模 Agentic 任务合成", "date": "2025-12-01", "lane": "后训练", "org": "deepseek-ai", "orgName": "DeepSeek", "releaseId": "deepseek-v3.2", "impact": "训练信号从对话搬到真实任务环境，智能体能力成为主线，也为后来的沙箱集群铺路。", "magnitude": "IMO / IOI 2025 金牌表现（Speciale）", "mechanism": "自动合成大批可验证的智能体任务环境，配合可扩展 RL 框架做大规模训练。", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "visual-causal-flow", "label": "VCF", "name": "Visual Causal Flow", "date": "2026-01-27", "lane": "多模态", "org": "deepseek-ai", "orgName": "DeepSeek", "releaseId": "deepseek-ocr-2", "impact": "视觉编码开始向人类的阅读方式靠拢，文档理解与后续多模态模型共享同一条路线。", "magnitude": "动态分辨率 + 光学压缩", "mechanism": "让视觉编码按因果阅读顺序推进，而不是按固定栅格扫描，并保留动态分辨率。", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "csa-hca", "label": "CSA + HCA", "name": "混合压缩注意力 + mHC", "date": "2026-04-22", "lane": "架构", "org": "deepseek-ai", "orgName": "DeepSeek", "releaseId": "deepseek-v4-flash", "impact": "百万上下文第一次不需要为长度付出线性代价，KV 与 FLOPs 同时掉一个数量级。", "magnitude": "1M 上下文：27% FLOPs、10% KV", "mechanism": "压缩稀疏注意力与重压缩注意力并行，配合流形约束超连接（mHC）与 Muon 优化器。", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "dspark", "label": "DSpark", "name": "模块化投机解码", "date": "2026-06-27", "lane": "推理", "org": "deepseek-ai", "orgName": "DeepSeek", "releaseId": "deepseek-v4-flash-dspark", "impact": "解码加速从「重训模型」变成「换启动参数」，推理优化与模型发布彻底解耦。", "magnitude": "vLLM / SGLang 单参数开启", "mechanism": "不改权重，只在同一 checkpoint 上挂一个投机解码模块，启动时加一个参数即可启用。", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "ced", "label": "CED", "name": "因果编码器-解码器架构", "date": "2026-09-10", "lane": "架构", "org": "deepseek-ai", "orgName": "DeepSeek", "releaseId": "deepseek-v4.1-flash", "impact": "预填充与解码被解耦成两种工作负载，各自只激活需要的部分，prefill 计算量约减半。", "magnitude": "prefill 8B / decode 16B 激活", "mechanism": "把网络拆成 20 层编码器与 20 层解码器，decoder 的全局 KV 由第 20 层 hidden state 投影而来，prefill 只激活 8B、decode 只激活 16B。", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "csa2", "label": "CSA2", "name": "跨层共享 KV 与索引", "date": "2026-09-10", "lane": "架构", "org": "deepseek-ai", "orgName": "DeepSeek", "releaseId": "deepseek-v4.1-flash", "impact": "索引成本从随上下文线性增长变成常数，跨层复用让深层的稀疏注意力几乎免费。", "magnitude": "indexer 成本线性 → 常数", "mechanism": "跨层共享 main KV 与 indexer K，复用 Top-K 索引；每层可选 Full / Reindex / Reuse 三种模式，深层由分层稀疏索引器接管。", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "fp4-kv", "label": "FP4 KV", "name": "FP4 主 KV 缓存", "date": "2026-09-10", "lane": "推理", "org": "deepseek-ai", "orgName": "DeepSeek", "releaseId": "deepseek-v4.1-flash", "impact": "KV 存储再砍一半，且精度损失可控，长上下文推理的显存天花板被再次抬高。", "magnitude": "main KV 存储近减半", "mechanism": "main KV 用 MXFP4（E2M1）加每 16 通道一个 E4M3 scale，省掉二级 global scale，在 RoPE 之后量化；SWA 部分仍保持 FP8。", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "swb-replay", "label": "SWB Replay", "name": "有界回放的滑窗 KV", "date": "2026-09-10", "lane": "推理", "org": "deepseek-ai", "orgName": "DeepSeek", "releaseId": "deepseek-v4.1-flash", "impact": "用便宜的宿主内存承接长上下文的冷 KV，整机显存不再是唯一约束。", "magnitude": "全局 KV 890 bytes/token，约 V4-Flash 的 1/4", "mechanism": "滑窗 KV 放进各机器 10% 的 host DRAM 组成分布式池，未命中时只回放最近一个窗口的 token。", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "engram", "label": "Engram", "name": "条件记忆模块", "date": "2026-09-10", "lane": "架构", "org": "deepseek-ai", "orgName": "DeepSeek", "releaseId": "deepseek-v4.1-flash", "impact": "把「记住常见片段」这件事从注意力里外包出去，注意力可以专注在真正需要推理的位置。", "magnitude": "196B 参数的条件记忆", "mechanism": "用 n-gram（2/3/4）查表式的条件记忆挂在第 1 与第 14 层，8 个 hash head，196B 参数以 FP8 存放。", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "single-pass-mhc", "label": "mHC", "name": "单遍残差混流", "date": "2026-09-10", "lane": "训练", "org": "deepseek-ai", "orgName": "DeepSeek", "releaseId": "deepseek-v4.1-flash", "impact": "激活值的内存流量减半，长序列训练的带宽瓶颈被松开，训练与推理都直接受益。", "magnitude": "残差读写 (3n+2)d → (2n+2)d", "mechanism": "把残差更新、输入混合与系数预测融合成一次单遍计算，残差读写降到理论下界 (2n+2)d。", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "effort-dial", "label": "Effort 1-100", "name": "可控推理强度", "date": "2026-09-10", "lane": "推理", "org": "deepseek-ai", "orgName": "DeepSeek", "releaseId": "deepseek-v4.1-flash", "impact": "同一模型按任务价值分配算力：effort 从 25 提到 100，8 项推理基准均分从 67.1% 涨到 76.3%。", "magnitude": "effort 25→100：67.1% → 76.3%", "mechanism": "推理强度做成 1 到 100 的标量，公开 API 对应 low=50 / high=75 / max=100。", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "dsec", "label": "DSec", "name": "百万级并发沙箱", "date": "2026-09-10", "lane": "后训练", "org": "deepseek-ai", "orgName": "DeepSeek", "releaseId": "deepseek-v4.1-flash", "impact": "Agentic RL 的数据瓶颈从「任务不够」变成「环境不够」，这一步把环境供给做成了基础设施。", "magnitude": "单机密度 1k → 2,500+", "mechanism": "自研安全沙箱，用 AppArmor 与 eBPF 做隔离，单机容器密度从 1,000 提到 2,500 以上。", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "glm-chatglm-6b", "label": "RLHF", "name": "以 62 亿参数配合 INT4 量化", "date": "2023-03-13", "lane": "推理", "org": "zai-org", "orgName": "智谱 GLM", "releaseId": "chatglm-6b", "impact": "智谱开源系列的第一代中英双语对话模型，62 亿参数、GLM 架构，可在消费级显卡本地部署，是整个 ChatGLM/GLM 开源谱系的起点。", "magnitude": "总参 62 亿", "mechanism": "以 62 亿参数配合 INT4 量化，把中文对话大模型拉到 6GB 显存的消费级显卡上运行。", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "glm-chatglm2-6b", "label": "GLM·2-6B", "name": "上下文长度由初代的 2K 扩展至 32K", "date": "2023-06-24", "lane": "训练", "org": "zai-org", "orgName": "智谱 GLM", "releaseId": "chatglm2-6b", "impact": "ChatGLM-6B 的第二代版本，借 FlashAttention 与 Multi-Query Attention 把上下文从 2K 提到 32K、推理提速 42%，同尺寸开源模型中竞争力显著增强。", "magnitude": "上下文 32K", "mechanism": "上下文长度由初代的 2K 扩展至 32K（对话 8K 训练），同时推理速度提升 42%、INT4 下 6G 显存支持对话长度由 1K 提升到 8K。", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "glm-codegeex2-6b", "label": "INT4", "name": "仅 60 亿参数即在 HumanEval-X", "date": "2023-07-19", "lane": "训练", "org": "zai-org", "orgName": "智谱 GLM", "releaseId": "codegeex2-6b", "impact": "CodeGeeX2-6B 是基于 ChatGLM2-6B 架构、经 600B 代码 token 预训练的多语言代码生成模型，Python HumanEval-X Pass@1 达 35.9%，用 6B 参数超过 15B 的 StarCoder-15B。", "magnitude": "上下文 8K", "mechanism": "仅 60 亿参数即在 HumanEval-X 上超越 150 亿参数的 StarCoder-15B（Python Pass@1 35.9%）。", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "glm-chatglm3-6b", "label": "Prompt", "name": "原生支持工具调用", "date": "2023-10-25", "lane": "后训练", "org": "zai-org", "orgName": "智谱 GLM", "releaseId": "chatglm3-6b", "impact": "ChatGLM 系列第三代的对话模型，最大变化是采用全新 Prompt 格式并原生支持 Function Call、Code Interpreter 与 Agent 场景，同时开源 Base 与 32K 长文本版本。", "magnitude": "总参 6B", "mechanism": "原生支持工具调用（Function Call）与代码执行，从纯对话模型转向可执行任务的 Agent 形态。", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "glm-cogvlm-chat-hf", "label": "MLP", "name": "视觉专家模块", "date": "2023-11-16", "lane": "多模态", "org": "zai-org", "orgName": "智谱 GLM", "releaseId": "cogvlm-chat-hf", "impact": "CogVLM-17B 是开源视觉语言模型，由 10B 视觉参数与 7B 语言参数组成，在 10 个跨模态基准上自称 SOTA，并在 VQAv2/OKVQA/TextVQA/COCO captioning 上排名第二。", "magnitude": "总参 17B", "mechanism": "视觉专家模块（Visual Expert）：在预训练语言模型的每一层引入视觉专家，使 VLM 效果超越或持平 PaLI-X 55B（自述）。", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "glm-cogagent-chat-hf", "label": "GUI·CogAgent-18B", "name": "首个把 VLM 用作 GUI Agent", "date": "2023-12-15", "lane": "多模态", "org": "zai-org", "orgName": "智谱 GLM", "releaseId": "cogagent-chat-hf", "impact": "CogAgent-18B 是在 CogVLM 基础上改进的视觉语言模型（11B 视觉 + 7B 语言），主打 GUI Agent、高分辨率图像理解与 OCR 增强，是 cogagent-chat 版本权重。", "magnitude": "总参 18B", "mechanism": "首个把 VLM 用作 GUI Agent：对任意 GUI 截图直接返回任务计划、下一步动作与操作坐标，并支持 1120x1120 超高清输入。", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "glm-cogvlm2-llama3-chat-19b", "label": "GPT", "name": "图文文档理解大幅跃升", "date": "2024-05-16", "lane": "多模态", "org": "zai-org", "orgName": "智谱 GLM", "releaseId": "cogvlm2-llama3-chat-19b", "impact": "CogVLM2 系列基于 Meta-Llama-3-8B-Instruct 构建的 19B 视觉语言模型，支持 8K 文本与 1344×1344 图像输入，DocVQA 92.3、TextVQA 84.2。", "magnitude": "上下文 8K", "mechanism": "图文文档理解大幅跃升：DocVQA 92.3 超过 GPT-4V(88.4)、OCRBench 756 超过 GPT-4V(656)，且全程不使用外部 OCR 工具。", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "glm-glm-4-9b-chat", "label": "BFCL", "name": "把 128K 长上下文、26 种多语言与 F", "date": "2024-06-04", "lane": "推理", "org": "zai-org", "orgName": "智谱 GLM", "releaseId": "glm-4-9b-chat", "impact": "GLM-4 系列的开源对话模型 GLM-4-9B-Chat，支持 128K 上下文、26 种语言、网页浏览与工具调用，是同代 9B 级开源模型的主力版本。", "magnitude": "上下文 128K", "mechanism": "把 128K 长上下文、26 种多语言与 Function Call 工具调用同时下放到 9B 开源模型（BFCL Overall Acc. 81.00，逼近 gpt-4-turbo 的 81.24）。", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "glm-glm-4v-9b", "label": "AI2D", "name": "高分辨率文字识别反超闭源旗舰", "date": "2024-06-04", "lane": "多模态", "org": "zai-org", "orgName": "智谱 GLM", "releaseId": "glm-4v-9b", "impact": "GLM-4 系列的开源多模态版本，具备 1120×1120 高分辨率中英双语多轮对话能力，卡内表格自述在中英文综合、感知推理、文字识别、图表理解上超越 GPT-4-turbo-2024-04-09、Gemini 1.0 Pro、Qwen-VL-Max 与 Claude 3 Opus。", "magnitude": "上下文 8K", "mechanism": "高分辨率文字识别反超闭源旗舰：OCRBench 786 为卡内同表最高（GPT-4o 736、GPT-4v 656），MMBench-EN 81.1 接近 GPT-4o 的 83.4。", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "glm-cogvideox-2b", "label": "torchao", "name": "把文生视频门槛压到消费级", "date": "2024-08-05", "lane": "多模态", "org": "zai-org", "orgName": "智谱 GLM", "releaseId": "cogvideox-2b", "impact": "CogVideoX 开源文生视频模型的入门版本（2B，FP16 训练），输出 720×480、6 秒、8 fps 视频，主打低成本推理与二次开发。", "magnitude": "总参 2B", "mechanism": "把文生视频门槛压到消费级：diffusers FP16 最低 4GB、INT8(torchao) 3.6GB 显存即可推理（推理速度 A100 约 90 秒/50 步）。", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "glm-cogvideox-5b", "label": "BF16·CogVideoX-5B", "name": "用 3d_rope_pos_embed 替换", "date": "2024-08-17", "lane": "多模态", "org": "zai-org", "orgName": "智谱 GLM", "releaseId": "cogvideox-5b", "impact": "CogVideoX 开源文生视频模型的较大规模版本（5B，BF16 训练），相比 2B 版本提供更高的视频生成质量与视觉效果。", "magnitude": "总参 5B", "mechanism": "用 3d_rope_pos_embed 替换 3d_sincos_pos_embed 以提升生成质量，推理显存最低 5GB（多卡 15GB，A100 约 180 秒/50 步）。", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "glm-glm-4-voice-9b", "label": "GLM·4-Voice-9B", "name": "端到端语音对话并可按指令改变情感、语调、语速", "date": "2024-10-24", "lane": "多模态", "org": "zai-org", "orgName": "智谱 GLM", "releaseId": "glm-4-voice-9b", "impact": "GLM-4-Voice 端到端语音模型的 LLM 部分：在 GLM-4-9B 基础上做语音模态预训练与对齐，可直接理解和生成中英文离散化语音并实时对话。", "magnitude": "总参 9B", "mechanism": "端到端语音对话并可按指令改变情感、语调、语速与方言，无需级联 ASR+TTS。", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "glm-cogvideox1-5-5b", "label": "BF16·CogVideoX1.5", "name": "分辨率从 720×480 提升到 1360×", "date": "2024-11-02", "lane": "多模态", "org": "zai-org", "orgName": "智谱 GLM", "releaseId": "cogvideox1-5-5b", "impact": "CogVideoX 的 1.5 代 5B 文生/图生视频模型（模型卡标注发布日 2024-11-08），分辨率升到 1360×768、帧率 16 fps、支持 5 秒或 10 秒视频。", "magnitude": "总参 5B", "mechanism": "分辨率从 720×480 提升到 1360×768、帧率 8→16 fps，视频时长由 6 秒扩展到 5/10 秒（A100 约 1000 秒/50 步）。", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "glm-glm-4-32b-0414", "label": "SWE·4-32B-0414", "name": "32B 稠密模型在 IFEval", "date": "2025-04-07", "lane": "训练", "org": "zai-org", "orgName": "智谱 GLM", "releaseId": "glm-4-32b-0414", "impact": "GLM-4-0414 系列的主力 32B 稠密模型（2025-04 发布，MIT 许可），基座 GLM-4-32B-Base-0414 用 15T 数据预训练，卡内自述在多项基准上对标 GPT-4o 与 DeepSeek-V3-0324(671B)。", "magnitude": "总参 32B", "mechanism": "32B 稠密模型在 IFEval(87.6)、BFCL-v3(69.6)、TAU-Bench Retail(68.7)、SimpleQA(88.1) 上追平乃至超过 GPT-4o-1120 与 671B 的 DeepSeek-V3-0324，并以 MIT 许可开源。", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "glm-glm-z1-32b-0414", "label": "RL·Z1-32B-0414", "name": "在 32B 开源规模上做出对标 DeepSe", "date": "2025-04-08", "lane": "训练", "org": "zai-org", "orgName": "智谱 GLM", "releaseId": "glm-z1-32b-0414", "impact": "基于 GLM-4-32B-0414 经冷启动与扩展强化学习打造的 32B 深度思考推理模型。", "magnitude": "上下文 32K", "mechanism": "在 32B 开源规模上做出对标 DeepSeek V3/R1 的深度推理能力，数学与复杂任务较底座显著提升。", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "glm-glm-4-1v-9b-thinking", "label": "RL·4.1V", "name": "首个以「思考范式 + 强化学习」训练的 GL", "date": "2025-06-28", "lane": "多模态", "org": "zai-org", "orgName": "智谱 GLM", "releaseId": "glm-4-1v-9b-thinking", "impact": "基于 GLM-4-9B-0414 的 9B 视觉语言推理模型，在 10B 级 VLM 中取得 SOTA。", "magnitude": "上下文 64K", "mechanism": "首个以「思考范式 + 强化学习」训练的 GLM-V 模型，在 28 项基准中 23 项为 10B 级最佳、18 项超越 Qwen-2.5-VL-72B。", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "glm-glm-4-5", "label": "A32B", "name": "以 355B-A32B 规模在 12 项标准", "date": "2025-07-20", "lane": "推理", "org": "zai-org", "orgName": "智谱 GLM", "releaseId": "glm-4-5", "impact": "面向智能体的 355B-A32B MoE 基础模型，统一推理、编码与智能体能力，支持思考/非思考双模式。", "magnitude": "上下文 128K", "mechanism": "以 355B-A32B 规模在 12 项标准基准上取得 63.2 的综合分，位列所有专有与开源模型第 3。", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "glm-glm-4-5-air", "label": "A12B", "name": "以 106B-A12B 的紧凑设计拿到 59", "date": "2025-07-20", "lane": "推理", "org": "zai-org", "orgName": "智谱 GLM", "releaseId": "glm-4-5-air", "impact": "GLM-4.5 系列中更紧凑高效的 106B-A12B MoE 版本，兼顾性能与部署效率。", "magnitude": "上下文 128K", "mechanism": "以 106B-A12B 的紧凑设计拿到 59.8 的综合分，接近 355B 旗舰而推理成本大幅降低。", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "glm-glm-4-5v", "label": "GLM·4.5V", "name": "在 42 项公开视觉语言基准上取得同规模 S", "date": "2025-08-10", "lane": "多模态", "org": "zai-org", "orgName": "智谱 GLM", "releaseId": "glm-4-5v", "impact": "以 GLM-4.5-Air（106B-12B active）为底座的旗舰视觉语言模型，覆盖图像、视频、文档理解与 GUI 操作。", "magnitude": "激活 12B", "mechanism": "在 42 项公开视觉语言基准上取得同规模 SOTA，并把 Grounding 与 GUI Agent 能力做进模型。", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "glm-glm-4-6", "label": "TIR", "name": "上下文从 128K 扩到 200K 以支撑更", "date": "2025-09-29", "lane": "推理", "org": "zai-org", "orgName": "智谱 GLM", "releaseId": "glm-4-6", "impact": "在 GLM-4.5 基础上把上下文扩到 200K 并全面强化编码、推理、工具与写作的迭代版本。", "magnitude": "上下文 200K", "mechanism": "上下文从 128K 扩到 200K 以支撑更复杂 Agent 任务，在 8 项公开基准上对 GLM-4.5 全面提分。", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "glm-glm-4-6v-flash", "label": "HTML", "name": "首次在 GLM-V 系列引入原生 Funct", "date": "2025-12-07", "lane": "多模态", "org": "zai-org", "orgName": "智谱 GLM", "releaseId": "glm-4-6v-flash", "impact": "GLM-4.6V 系列的 9B 轻量多模态模型，面向本地部署与低延迟场景。", "magnitude": "上下文 128K", "mechanism": "首次在 GLM-V 系列引入原生 Function Calling，把「视觉感知」直接接到「可执行动作」。", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "glm-autoglm-phone-9b", "label": "GUI·AutoGLM", "name": "把视觉语言模型直接接进真实手机 GUI", "date": "2025-12-08", "lane": "多模态", "org": "zai-org", "orgName": "智谱 GLM", "releaseId": "autoglm-phone-9b", "impact": "AutoGLM 手机智能助手框架中的开源模型，理解手机屏幕并自动完成跨 App 操作任务。", "magnitude": "总参 9B", "mechanism": "把视觉语言模型直接接进真实手机 GUI：自然语言下达任务即可自动解析意图、规划并执行操作。", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "glm-glm-asr-nano-2512", "label": "V3", "name": "以 1.5B 规模取得同类开源模型最低平均错", "date": "2025-12-09", "lane": "多模态", "org": "zai-org", "orgName": "智谱 GLM", "releaseId": "glm-asr-nano-2512", "impact": "1.5B 的开源语音识别模型，重点优化方言与低音量场景。", "magnitude": "总参 1.5B", "mechanism": "以 1.5B 规模取得同类开源模型最低平均错误率 4.10，并在多基准上超过 OpenAI Whisper V3。", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "glm-glm-4-7", "label": "SWE·4.7", "name": "SWE-bench Verified 73.", "date": "2025-12-22", "lane": "后训练", "org": "zai-org", "orgName": "智谱 GLM", "releaseId": "glm-4-7", "impact": "面向编码与 Agent 的旗舰模型，强化 UI/Vibe Coding、工具使用与复杂推理能力。", "magnitude": "见模型卡", "mechanism": "SWE-bench Verified 73.8%（+5.8）、Terminal Bench 2.0 41.0%（+16.5）、HLE w/ Tools 42.8%（+12.4）较 GLM-4.6 大幅提升，并新增 Preserved Thinking。", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "glm-glm-image", "label": "GLM·Image", "name": "CVTG-2K 词准确率 0.9116 / ", "date": "2026-01-08", "lane": "多模态", "org": "zai-org", "orgName": "智谱 GLM", "releaseId": "glm-image", "impact": "混合自回归 + 扩散解码器的图像生成模型，在文本渲染与知识密集型生成上优势明显。", "magnitude": "见模型卡", "mechanism": "CVTG-2K 词准确率 0.9116 / NED 0.9557、LongText-Bench ZH 0.979，文本渲染能力显著领先同表开源与闭源模型。", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "glm-glm-4-7-flash", "label": "A3B", "name": "以 30B-A3B 在 SWE-bench ", "date": "2026-01-19", "lane": "架构", "org": "zai-org", "orgName": "智谱 GLM", "releaseId": "glm-4-7-flash", "impact": "30B-A3B 的 MoE 轻量模型，主打 30B 级别性能与效率平衡。", "magnitude": "激活 3B", "mechanism": "以 30B-A3B 在 SWE-bench Verified 59.2、τ²-Bench 79.5 上大幅领先同量级 Qwen3-30B-A3B-Thinking-2507 与 GPT-OSS-20B。", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "glm-glm-ocr", "label": "MTP·OCR", "name": "以 0.9B 规模在 OmniDocBenc", "date": "2026-01-30", "lane": "多模态", "org": "zai-org", "orgName": "智谱 GLM", "releaseId": "glm-ocr", "impact": "0.9B 参数的多模态 OCR 模型，面向复杂文档解析与结构化信息抽取。", "magnitude": "总参 0.9B", "mechanism": "以 0.9B 规模在 OmniDocBench V1.5 取得 94.62 分排名第一，并可在 vLLM/SGLang/Ollama 上高并发部署。", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "glm-glm-5", "label": "DSA", "name": "从 GLM-4.5 的 355B-A32B ", "date": "2026-02-11", "lane": "训练", "org": "zai-org", "orgName": "智谱 GLM", "releaseId": "glm-5", "impact": "面向复杂系统工程与长程 Agent 任务的 744B-A40B 开源旗舰模型。", "magnitude": "上下文 200K", "mechanism": "从 GLM-4.5 的 355B-A32B 扩到 744B-A40B、预训练数据从 23T 增到 28.5T，并在推理/编码/Agent 上取得全球开源最佳，逼近前沿闭源模型。", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "glm-glm-5-1", "label": "SWE·5.1", "name": "在长程任务上可维持数百轮优化、数千次工具调用", "date": "2026-04-03", "lane": "后训练", "org": "zai-org", "orgName": "智谱 GLM", "releaseId": "glm-5-1", "impact": "面向 agentic 工程的下一代旗舰模型，重点提升长程任务中的持续优化能力。", "magnitude": "见模型卡", "mechanism": "在长程任务上可维持数百轮优化、数千次工具调用持续提分（SWE-Bench Pro 58.4 领先 GLM-5、Terminal-Bench 2.0 63.5 vs GLM-5 的 56.2）。", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "glm-glm-5-2", "label": "MTP·5.2", "name": "solid 1M 上下文 + IndexSh", "date": "2026-06-16", "lane": "训练", "org": "zai-org", "orgName": "智谱 GLM", "releaseId": "glm-5-2", "impact": "首个能在 1M token 上下文上稳定支撑长程任务的旗舰模型，并提出 IndexShare 稀疏注意力。", "magnitude": "上下文 1M", "mechanism": "solid 1M 上下文 + IndexShare 使 1M 长度下每 token FLOPs 降 2.9×，HLE 40.5、AIME 2026 99.2 大幅超越 GLM-5.1。", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "glm-glm-5-3", "label": "GLM·5.3", "name": "编码能力较 GLM-5.2 提升 50%", "date": "2026-08-25", "lane": "训练", "org": "zai-org", "orgName": "智谱 GLM", "releaseId": "glm-5-3", "impact": "与 GLM-5.2 共用基座、所有提升来自后训练的编码与长程任务强化版。", "magnitude": "上下文 1M", "mechanism": "编码能力较 GLM-5.2 提升 50%（自研 Z.ai Code Bench），CyberGym 84.5 居 SOTA，ExploitGym 2h/6h 达 105/130（GLM-5.2 为 29/39）。", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "glm-glm-5-3-flash", "label": "mHC", "name": "首次在 GLM 系列引入稀疏+线性混合注意力", "date": "2026-08-25", "lane": "多模态", "org": "zai-org", "orgName": "智谱 GLM", "releaseId": "glm-5-3-flash", "impact": "GLM-5 系列首个原生多模态模型，320B-A18B，主打低成本高性能。", "magnitude": "激活 18B", "mechanism": "首次在 GLM 系列引入稀疏+线性混合注意力与 mHC 超连接，320B-18B 以十分之一价格超过 GLM-5.2 并逼近 Claude Opus 4.8 的编码/Agent 水平。", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "kimi-moonlight-16b-a3b-instruct", "label": "RMS", "name": "首次证明 Muon 优化器可用于大规模 LL", "date": "2025-02-22", "lane": "训练", "org": "moonshotai", "orgName": "月之暗面 Kimi", "releaseId": "moonlight-16b-a3b-instruct", "impact": "用 Muon 优化器训练的 3B/16B MoE 模型，5.7T token 预训练，验证了 Muon 在大规模训练上的可扩展性。", "magnitude": "上下文 8K", "mechanism": "首次证明 Muon 优化器可用于大规模 LLM 训练：约 2x 样本效率、约 52% 训练 FLOPs 达到 AdamW 同等性能", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "kimi-kimi-vl-a3b-instruct", "label": "MLP", "name": "以 2.8B+0.4B 激活参数在 Scre", "date": "2025-04-09", "lane": "多模态", "org": "moonshotai", "orgName": "月之暗面 Kimi", "releaseId": "kimi-vl-a3b-instruct", "impact": "激活仅 2.8B 参数的高效 MoE VLM（16B 总参/3B 激活），128K 上下文，具备长视频/长文档与 OS Agent 能力。", "magnitude": "上下文 128K", "mechanism": "以 2.8B+0.4B 激活参数在 ScreenSpot-V2(92.8)、OSWorld(8.22) 等 GUI Agent 任务上超过 GPT-4o", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "kimi-kimi-dev-72b", "label": "SWE", "name": "SWE-bench Verified 达 6", "date": "2025-06-16", "lane": "后训练", "org": "moonshotai", "orgName": "月之暗面 Kimi", "releaseId": "kimi-dev-72b", "impact": "面向软件工程问题解决的编码 LLM（base_model 为 Qwen/Qwen2.5-72B）；模型卡未印参数量，故参数字段为 null。", "magnitude": "见模型卡", "mechanism": "SWE-bench Verified 达 60.4%，刷新开源模型 SOTA", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "kimi-kimi-vl-a3b-thinking-2506", "label": "think", "name": "多模态推理大幅提升并降低 20% 思维长度", "date": "2025-06-21", "lane": "多模态", "org": "moonshotai", "orgName": "月之暗面 Kimi", "releaseId": "kimi-vl-a3b-thinking-2506", "impact": "Kimi-VL-A3B-Thinking 的升级版：base_model 指向 Kimi-VL-A3B-Instruct（同代 16B 总参/3B 激活），上下文长度按卡内 VLLM 示例 max_model_len=131072(=128K) 记为 128K；本卡未印参数量表，参数与上下文档为同代推断。", "magnitude": "上下文 128K", "mechanism": "多模态推理大幅提升并降低 20% 思维长度：MathVision 36.8→56.9，单图支持 3.2M 像素（4x）", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "kimi-kimi-k2-base", "label": "MLA·K2-Base", "name": "MuonClip 优化器把 Muon 首次扩", "date": "2025-07-03", "lane": "训练", "org": "moonshotai", "orgName": "月之暗面 Kimi", "releaseId": "kimi-k2-base", "impact": "Kimi K2 的基座模型：1T 总参/32B 激活的 MoE，用 Muon 优化器在 15.5T token 上预训练且训练全程无失稳。", "magnitude": "上下文 128K", "mechanism": "MuonClip 优化器把 Muon 首次扩展到 1T 参数 MoE 规模，实现 15.5T token 零训练不稳定", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "kimi-kimi-k2-instruct", "label": "MLA·K2-Instruct", "name": "开源 1T 参数级 agentic 模型", "date": "2025-07-11", "lane": "训练", "org": "moonshotai", "orgName": "月之暗面 Kimi", "releaseId": "kimi-k2-instruct", "impact": "K2 的后训练版（非长思考的 reflex 级模型），1T 总参/32B 激活 MoE，专为工具使用与自主问题求解优化。", "magnitude": "上下文 128K", "mechanism": "开源 1T 参数级 agentic 模型：SWE-bench Verified 65.8%（agentic 单次）、Tau2/AceBench 等工具使用基准领先", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "kimi-kimi-linear-48b-a3b-instruct", "label": "KDA·Linear", "name": "KDA 混合线性注意力", "date": "2025-10-30", "lane": "推理", "org": "moonshotai", "orgName": "月之暗面 Kimi", "releaseId": "kimi-linear-48b-a3b-instruct", "impact": "混合线性注意力模型（KDA:MLA=3:1），5.7T token 训练，长上下文下相比全注意力最多 6x 解码吞吐提升。", "magnitude": "上下文 1M", "mechanism": "KDA 混合线性注意力：1M 上下文解码吞吐最高 6x 提升、KV cache 需求降低最多 75%，RULER(128k) 84.3", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "kimi-kimi-k2-thinking", "label": "QAT", "name": "QAT 原生 INT4 量化实现约 2x 生", "date": "2025-11-04", "lane": "推理", "org": "moonshotai", "orgName": "月之暗面 Kimi", "releaseId": "kimi-k2-thinking", "impact": "K2 的开源思考模型，边推理边动态调用工具，256K 上下文，采用 QAT 原生 INT4 量化。", "magnitude": "上下文 256K", "mechanism": "QAT 原生 INT4 量化实现约 2x 生成加速且性能无损，并在 HLE/BrowseComp 等基准刷新 SOTA", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "kimi-kimi-k2-5", "label": "INT4·K2.5", "name": "原生多模态 + Agent Swarm", "date": "2026-01-01", "lane": "多模态", "org": "moonshotai", "orgName": "月之暗面 Kimi", "releaseId": "kimi-k2-5", "impact": "在 Kimi-K2-Base 上以约 15 万亿图文混合 token 继续预训练得到的原生多模态 agentic 模型，含 Instant/Thinking 双模式与 Agent Swarm。", "magnitude": "上下文 256K", "mechanism": "原生多模态 + Agent Swarm：BrowseComp 由 60.6 提升到 74.9（上下文管理）、78.4（Agent Swarm）", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "kimi-kimi-k2-6", "label": "INT4·K2.6", "name": "Agent Swarm 横向扩展到 300 ", "date": "2026-04-14", "lane": "后训练", "org": "moonshotai", "orgName": "月之暗面 Kimi", "releaseId": "kimi-k2-6", "impact": "K2.5 的后继原生多模态 agentic 模型，强化长程编码、编码驱动设计与 300 子代理的 Agent Swarm 编排。", "magnitude": "上下文 256K", "mechanism": "Agent Swarm 横向扩展到 300 个子代理、4000 步协同，BrowseComp(Agent Swarm) 86.3、SWE-Bench Verified 80.2", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "kimi-kimi-k2-7-code", "label": "K2", "name": "相比 K2.6 思考 token 用量减少约", "date": "2026-06-11", "lane": "训练", "org": "moonshotai", "orgName": "月之暗面 Kimi", "releaseId": "kimi-k2-7-code", "impact": "基于 Kimi K2.6 打造的编码专用 agentic 模型，强化长程工程任务端到端完成能力，并提升 token 效率。", "magnitude": "上下文 256K", "mechanism": "相比 K2.6 思考 token 用量减少约 30%：Kimi Code Bench v2 50.9→62.0、MCP Mark Verified 72.8→81.1", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } }, { "id": "kimi-kimi-k3", "label": "KDA·K3", "name": "以 KDA+Attention Residu", "date": "2026-06-13", "lane": "推理", "org": "moonshotai", "orgName": "月之暗面 Kimi", "releaseId": "kimi-k3", "impact": "全球首个开源的 3T 级模型：2.8T 总参/104B 激活，基于 KDA + AttnRes + Stable LatentMoE，原生多模态，1M token 上下文（模型卡写作 1048576）。", "magnitude": "上下文 1M", "mechanism": "以 KDA+Attention Residuals+Stable LatentMoE（896 专家激活 16）把 MoE 稀疏度推高，整体扩展效率较 Kimi K2 提升约 2.5x", "review": { "count": 0, "lastReviewedAt": null, "by": null, "note": null } } ], "orgs": [ { "id": "deepseek-ai", "name": "DeepSeek", "short": "DeepSeek", "blurb": "从 7B 稠密模型到 671B 稀疏专家，MLA / GRPO / V4 系列的开源路线", "hue": 190 }, { "id": "zai-org", "name": "智谱 GLM", "short": "GLM", "blurb": "GLM 系列：从 ChatGLM 到 GLM-5.3，全栈开源（语言 / 视觉 / 代码 / 视频 / 语音）", "hue": 265 }, { "id": "moonshotai", "name": "月之暗面 Kimi", "short": "Kimi", "blurb": "Kimi 系列：Muon 优化器与超稀疏 MoE 路线的开源权重", "hue": 42 } ], "market": { "updatedAt": "2026-09-29", "quotes": [ { "device_id": "gb300_nvl72", "variant_id": "smc_gb300_nvl72", "vendor": "Supermicro", "price_cny": 127000000, "unit": "套", "date": "2026-09-29", "source_id": "src-market-20260929", "raw": "超微GB3001.27亿/套", "op": "gte" }, { "device_id": "b300_hgx_server", "variant_id": "smc_b300_4t", "vendor": "Supermicro", "price_cny": 16800000, "unit": "台", "date": "2026-09-29", "source_id": "src-market-20260929", "raw": "超微B300/4T 1680万/台" }, { "device_id": "b300_hgx_server", "variant_id": "smc_b300_3t", "vendor": "Supermicro", "price_cny": 16200000, "unit": "台", "date": "2026-09-29", "source_id": "src-market-20260929", "raw": "超微B3003T 1620万/台" }, { "device_id": "b300_hgx_server", "variant_id": "asus_b300_3t", "vendor": "ASUS", "price_cny": 16100000, "unit": "台", "date": "2026-09-29", "source_id": "src-market-20260929", "raw": "华硕B300/3T 1610万/台" }, { "device_id": "b300_hgx_server", "variant_id": "dell_b300_4t", "vendor": "Dell", "price_cny": 16600000, "unit": "台", "date": "2026-09-29", "source_id": "src-market-20260929", "raw": "戴尔B300/4T 1660万/台" }, { "device_id": "b300_hgx_server", "variant_id": "gigabyte_b300_3t", "vendor": "GIGABYTE", "price_cny": 16100000, "unit": "台", "date": "2026-09-29", "source_id": "src-market-20260929", "raw": "技嘉B300/3T 1610万/台" }, { "device_id": "h200_hgx_server", "variant_id": "smc_h200", "vendor": "Supermicro", "price_cny": 6150000, "unit": "台", "date": "2026-09-29", "source_id": "src-market-20260929", "raw": "超微H200 615万/台" }, { "device_id": "h200_hgx_server", "variant_id": "dell_h200", "vendor": "Dell", "price_cny": 6100000, "unit": "台", "date": "2026-09-29", "source_id": "src-market-20260929", "raw": "戴尔H200 610万/台" }, { "device_id": "h200_hgx_server", "variant_id": "gigabyte_h200", "vendor": "GIGABYTE", "price_cny": 6000000, "unit": "台", "date": "2026-09-29", "source_id": "src-market-20260929", "raw": "技嘉H200 600万/台" }, { "device_id": "h200_hgx_server", "variant_id": "asus_h200", "vendor": "ASUS", "price_cny": 6000000, "unit": "台", "date": "2026-09-29", "source_id": "src-market-20260929", "raw": "华硕H200 600万/台" }, { "device_id": "h200_hgx_server", "variant_id": "h200_module_8gpu", "vendor": "NVIDIA", "price_cny": 5100000, "unit": "台", "date": "2026-09-29", "source_id": "src-market-20260929", "raw": "H200模组 510万/台" }, { "device_id": "h100_hgx_server", "variant_id": "smc_h100", "vendor": "Supermicro", "price_cny": 3800000, "unit": "台", "date": "2026-09-29", "source_id": "src-market-20260929", "raw": "超微H100 380万/台" }, { "device_id": "h100_hgx_server", "variant_id": "gigabyte_h100", "vendor": "GIGABYTE", "price_cny": 3700000, "unit": "台", "date": "2026-09-29", "source_id": "src-market-20260929", "raw": "技嘉H100 370万/台" }, { "device_id": "rtx5090", "variant_id": "rtx5090_axial", "vendor": null, "price_cny": 40000, "unit": "张", "date": "2026-09-29", "source_id": "src-market-20260929", "raw": "5090风扇卡4万+/张", "op": "gte" }, { "device_id": "rtx5090", "variant_id": "rtx5090_blower", "vendor": null, "price_cny": 50000, "unit": "张", "date": "2026-09-29", "source_id": "src-market-20260929", "raw": "5090涡轮卡 5万+/张", "op": "gte" }, { "device_id": "rtx_pro6000_bw", "variant_id": "rtx_pro6000_bw_server", "vendor": null, "price_cny": 180000, "unit": "张", "date": "2026-09-29", "source_id": "src-market-20260929", "raw": "Pro6000服务器版 18万+/张", "op": "gte" } ] }, "reviews": { "updatedAt": "2026-11-03", "reviewer": "weidows", "note": "人工复核台账。本项目的数据由人工抓取整理，没有自动校验；每核对一条记录的字段来源就 +1，构建时由 scripts/build_data.py 合入 atlas.json 的 review 字段。count=0 表示这条只抓过、没做过二次核对。", "stats": { "total": 341, "reviewed": 0, "unreviewed": 341, "lastReviewedAt": null, "byKind": { "hardware": { "total": 174, "reviewed": 0, "unreviewed": 174 }, "server": { "total": 15, "reviewed": 0, "unreviewed": 15 }, "model": { "total": 84, "reviewed": 0, "unreviewed": 84 }, "breakthrough": { "total": 68, "reviewed": 0, "unreviewed": 68 } } } }, "counts": { "docxProducts": 154, "extraProducts": 20, "servers": 15, "models": 84, "breakthroughs": 68, "china": 8, "matched": 21 } }