第一项始终是图钉的来源。总体可信度是各份资料对上方所用开始和结束时间支持程度的加权平均;资料每比最新的一份旧 180 天,权重减半。
MiniMax[4]的文章日期为2026-06-01,开头是"MiniMax M3 is officially released today";其API平台发布说明将该模型列在2026年6月1日。
The open weights' model card: about 428B total and 23B activated parameters, 1M-token context, MiniMax[1][4] Sparse Attention in place of GQA with 9x prefill and 15x decode speedups at 1M, enabled/adaptive/disabled thinking, the minimax-community license, and 80.5 on SWE-bench Verified and 59 on SWE-bench Pro.
Independent evaluation: 29 on the Intelligence Index v4.3.2 (18th of 117), about 107 output tokens per second, $0.30 per 1 million input and $1.20 output, $0.51 per Intelligence Index task, 1M context, text/image/video input, 428B total and 23B active.[1]
MiniMax's[1] API price table: MiniMax-M3 up to 512k input at $0.60 (now $0.30 with the permanent 50% off) input, $2.40 ($1.20) output and $0.12 ($0.06) prompt-cache read per 1 million tokens, and above 512k $1.20 ($0.60), $4.80 ($2.40) and $0.24 ($0.12).
The MSA paper: blockwise sparse attention on top of GQA with a lightweight index branch, 28.4x less per-token attention compute at 1M context and 14.2x prefill and 7.6x decode speedups on H800 GPUs on a 109B test model; says MiniMax[1][4]-M3 is its production model.
有遗漏或错误吗?用你自己的话说明:能佐证此图钉的链接、不同的开始或结束日期及理由,或缺失、有误的信息。AI 会对照此图钉的来源进行核实,搜索更好的来源,并添加任何支持你说法的页面。图钉自身的来源仍然最重要。AI 也会查看图片:显示的是别的东西或显示效果差的图片会被移到后面或替换。
原标题: MiniMax-M3 Released
| 基准 | MiniMax M3 | Claude Opus 4.7 | GPT 5.5 | Gemini 3.1 Pro |
|---|---|---|---|---|
| SWE-Bench Pro | 59.0 | 64.3 | 58.6 | 54.2 |
| Terminal Bench 2.1 | 66.0 | 66.1 | 78.2 | 70.0 |
| VIBE V2 | 50.1 | 55.8 | 50.5 | 28 |
| SVG-Bench | 63.7 | 62.3 | 58.2 | 59.2 |
| KernelBench Hard | 28.8 | 30.7 | 20.9 | 18.6 |
| BrowseComp | 83.5 | 79.3 | 84.4 | 85.9 |
| GDPval评分细则 | 74.7 | 79.8 | 80.6 | 57.8 |
| BankerToolBench | 76.1 | 81.3 | 70.0 | 67.0 |
| MCP Atlas | 74.2 | 77 | 75.3 | 69.2 |
| OSWorld-verified | 75.2 | 82.8 | 78.7 | 76.2 |