读者上传的照片,与图钉带来源的媒体分开显示。你可以为吃到或看到的内容评分,在餐厅图钉上还能直接在照片上标记菜品。
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[3] 的文章《MiniMax M2.5: Built for Real-World Productivity》日期为 2026.2.12,开头为“Today we’re introducing our latest model, MiniMax-M2.5”;API 发布说明将 M2.5 系列列在 2026 年 2 月。
The open-weights model card: 229B parameters, Modified-MIT licence, the M2.5-Lightning and standard prices and speeds, and the results list (80.2% SWE-Bench Verified, 51.3% Multi-SWE-Bench, 76.3% BrowseComp with context management, 86.3 AIME25, 85.2 GPQA-Diamond).[1]
The current API price table: MiniMax[1]-M2.5 at $0.30 per 1 million input tokens, $1.20 output, $0.03 cache read and $0.375 cache write; MiniMax-M2.5-highspeed at $0.60 and $2.40; M2.1 and M2 at the same $0.30 and $1.20.
Independent profile: 23 (estimated) on the Intelligence Index v4.3.2, about 68 output tokens per second, a 205k-token context window, 230B total and 10B active parameters, $0.30 per 1 million input and $1.20 output.[1]
The series paper that reviews M2 through M2.7: 229.9B total and 9.8B active parameters, the Forge agent RL system, and agent-driven data pipelines for agentic coding and office work.[1]
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-M2.5 Released
| 基准 | M2.5 | M2.1 | Claude Opus 4.5 | Claude Opus 4.6 | Gemini 3 Pro | GPT-5.2 |
|---|---|---|---|---|---|---|
| SWE-Bench Verified | 80.2 | 74 | 80.9 | 80.8 | 78 | 80 |
| SWE-Bench Pro | 55.4 | 49.7 | 56.9 | 55.4 | 54.1 | 55.6 |
| SWE-Bench Multilingual | 74.1 | 71.9 | 77.5 | 77.8 | 65 | 72 |
| Multi-SWE-Bench | 51.3 | 47.2 | 50 | 50.3 | 42.7 | — |
| Terminal Bench 2 | 51.7 | 47.9 | 53.4 | 55.1 | 54 | 54 |
| VIBE-Pro (平均) | 54.2 | 42.4 | 55.2 | 55.6 | 36.9 | — |
| BrowseComp (含上下文管理) | 76.3 | 62 | 67.8 | 84 | 59.2 | 65.8 |
| BFCL 多轮 | 76.8 | 37.4 | 68 | 63.3 | 61 | — |
| MEWC (Excel) | 74.4 | 55.6 | 82.1 | 89.8 | 78.7 | 41.3 |
| GDPval-MM | 59 | 24.6 | 61.1 | 73.5 | 28.1 | 54.5 |