MI

MiniMax-M2.7

Reasoning modelMiniMax MMiniMax M2.7

MiniMax-M2.7

Release date: 2026-03-18Updated: 2026-04-12 15:22:19.6589,684
Parameters
229B
Context length
200K
Chinese support
Supported
Reasoning ability

MiniMax-M2.7 是 MiniMax 于 2026 年 3 月发布的推理大模型,采用 MoE(混合专家)架构,总参数量 2300 亿、激活参数 100 亿,支持 200K tokens 超长上下文。M2.7 最大的技术亮点是首次将模型引入自身训练循环:基于 M2.7 构建的强化学习 Harness 驱动了实验监控、日志排查、代码修复与评测循环,模型可承担相关研发工作流约 30–50% 的工作量。在软件工程方向,M2.7 在 SWE-Pro 评测上得分 56.22%,接近 GPT-5.3-Codex;在 SWE Multilingual 上达到 76.5;在专业办公场景中,GDPval-AA ELO 得分 1500,位列全球第四;工具调用基准 Toolathon 正确率 46.3%。M2.7 目前已通过 MiniMax Agent(agent.minimaxi.com)和 API 平台(platform.minimaxi.com)全量上线,API 输入定价 $0.30/1M tokens,输出 $1.20/1M tokens,不开源。

Data sourced primarily from official releases (GitHub, Hugging Face, papers), then benchmark leaderboards, then third-party evaluators. Learn about our data methodology

MiniMax-M2.7

Model basics

Reasoning traces
Supported
Thinking modes
Thinking Mode (Default)Standard Mode
Context length
200K tokens
Max output length
200K tokens
Model type
Reasoning model
Modality (in / out)
No data
Release date
2026-03-18
Model file size
未知
MoE architecture
Yes
Total params / Active params
229B / 10B
Knowledge cutoff
No data
MiniMax-M2.7

Open source & experience

MiniMax-M2.7

Official resources

MiniMax-M2.7

API details

API speed
3/5
💡Default unit: $/1M tokens. If vendors use other units, follow their published pricing.
Standard
TypeConditionInputOutput
Text-$0.300/ 1M$1.20/ 1M
Cache PricingPrompt Cache
TypeTTLWriteRead
Text5m$0.375/ 1M$0.060/ 1M
MiniMax-M2.7

Benchmark Results

MiniMax-M2.7 currently shows benchmark results led by IF Bench (5 / 29, score 76), Claw Bench (5 / 29, score 91.70), GPQA Diamond (39 / 179, score 87). This page also consolidates core specs, context limits, and API pricing so you can evaluate the model from benchmark results and deployment constraints together.

Thinking
Tool usage

General Knowledge

2 evaluations
Benchmark / mode
Score
Rank/total
GPQA Diamond
Thinking Mode
87
39 / 179
HLE
Thinking Mode
28
84 / 159

Coding and Software Engineer

1 evaluations
Benchmark / mode
Score
Rank/total
SWE-Bench Pro - Public
Thinking ModeTools
56.20
17 / 44

Agent Level Benchmark

2 evaluations
Benchmark / mode
Score
Rank/total
τ²-Bench - Telecom
Thinking ModeTools
85
24 / 35
Terminal Bench Hard
Thinking ModeTools
39
5 / 13

Instruction Following

1 evaluations
Benchmark / mode
Score
Rank/total
IF Bench
Thinking ModeTools
76
5 / 29

Productivity Knowledge

1 evaluations
Benchmark / mode
Score
Rank/total
GDPval-AA
Thinking Mode
50
13 / 21

Long Context

1 evaluations
Benchmark / mode
Score
Rank/total
AA-LCR
Thinking ModeTools
69
4 / 13

Claw-style Agent Evaluation

2 evaluations
Benchmark / mode
Score
Rank/total
Claw Bench
Thinking ModeTools
91.70
5 / 29
Pinch Bench
Thinking ModeTools
87.10
9 / 37

Compare with other models

MiniMax-M2.7

Publisher

MiniMax-M2.7

Model Overview

MiniMax-M2.7 是 MiniMax 于 2026 年 3 月发布的推理大模型,采用 MoE(混合专家)架构,总参数量 2300 亿、激活参数 100 亿,支持 200K tokens 超长上下文。M2.7 最大的技术亮点是首次将模型引入自身训练循环:基于 M2.7 构建的强化学习 Harness 驱动了实验监控、日志排查、代码修复与评测循环,模型可承担相关研发工作流约 30–50% 的工作量。在软件工程方向,M2.7 在 SWE-Pro 评测上得分 56.22%,接近 GPT-5.3-Codex;在 SWE Multilingual 上达到 76.5;在专业办公场景中,GDPval-AA ELO 得分 1500,位列全球第四;工具调用基准 Toolathon 正确率 46.3%。M2.7 目前已通过 MiniMax Agent(agent.minimaxi.com)和 API 平台(platform.minimaxi.com)全量上线,API 输入定价 $0.30/1M tokens,输出 $1.20/1M tokens,不开源。

DataLearner on WeChat

Follow DataLearner on WeChat for AI model updates and research notes.

DataLearner WeChat QR code