MiniMax M3
MiniMax M3 于 2026 年 6 月 1 日正式发布,采用 MoE 架构(428B 总参数,23B 每 token 激活),并引入自研稀疏注意力架构 MSA(MiniMax Sparse Attention)。在 1M token 上下文下解码速度较上代提升 15.6 倍,支持最高 1M tokens 超长上下文与原生多模态(图片、视频输入及桌面操作)。SWE-Bench Pro 达到 59.0%(超越 GPT-5.5 和 Gemini 3.1 Pro),BrowseComp 得分 83.5(超越 Opus 4.7)。模型权重已在 HuggingFace 和 GitHub 开源。
Data sourced primarily from official releases (GitHub, Hugging Face, papers), then benchmark leaderboards, then third-party evaluators. Learn about our data methodology
Model basics
Open source & experience
Official resources
API details
| Type | Condition | Input | Output |
|---|---|---|---|
| Text | - | ¥2.10/ 1M | ¥8.40/ 1M |
| Text | Context <= 524288 | ¥4.20/ 1M | ¥16.80/ 1M |
| Type | Condition | Input | Output |
|---|---|---|---|
| Text | - | ¥3.15/ 1M | ¥12.60/ 1M |
| Text | Context <= 524288 | ¥6.30/ 1M | ¥25.20/ 1M |
| Type | TTL | Write | Read |
|---|---|---|---|
| Text | - | - | ¥0.420/ 1M |
Benchmark Results
MiniMax M3 currently shows benchmark results led by BrowseComp (10 / 50, score 83.50), SWE-Bench Pro - Public (10 / 49, score 59), LiveBench (40 / 115, score 70.02). This page also consolidates core specs, context limits, and API pricing so you can evaluate the model from benchmark results and deployment constraints together.
Coding and Software Engineer
1 evaluationsAI Agent - Tool Usage
2 evaluationsCompare with other models
No curated comparisons for this model yet.
Want a custom combination? Open the compare tool
Publisher
Model Overview
MiniMax M3 于 2026 年 6 月 1 日正式发布,采用 MoE 架构(428B 总参数,23B 每 token 激活),并引入自研稀疏注意力架构 MSA(MiniMax Sparse Attention)。在 1M token 上下文下解码速度较上代提升 15.6 倍,支持最高 1M tokens 超长上下文与原生多模态(图片、视频输入及桌面操作)。SWE-Bench Pro 达到 59.0%(超越 GPT-5.5 和 Gemini 3.1 Pro),BrowseComp 得分 83.5(超越 Opus 4.7)。模型权重已在 HuggingFace 和 GitHub 开源。
DataLearner on WeChat
Follow DataLearner on WeChat for AI model updates and research notes.
