MiniMaxAIOpen Source

MiniMax-M2.7

Name: MiniMax-M2.7
Author: MiniMaxAI

Compare this model

A high-performance model developed by MiniMax. This improved version of 2.5 achieves higher performance.

Share:X Hatena LINE

Parameters

Undisclosed

Context Window

License

https://github.com/MiniMax-AI/MiniMax-M2.7/blob/main/LICENSE

Release Date

2026-03-18

Japanese Language Capability

✅High-Quality JP

Multilingual model with strong Japanese language processing capabilities.

API Pricing

Input Price (per 1M tokens)

$0.3

Output Price (per 1M tokens)

Billing Mode: standard

Strengths

Weaknesses

Use Cases

Deep Analysis

Intelligence Index

Artificial Analysis; open-weight tier-1 coding specialist

SWE-bench Verified

78%

Resolving real GitHub issues with human verification

SWE-Pro

56.22%

Professional software-engineering eval

Terminal-Bench 2

57.0%

Agentic terminal / tool-use tasks

Input Price

$0.30/1M

Output $1.20/1M; cached input $0.06/1M; high-speed ~$0.60/$2.40

Context Window

205K tokens

Text-only; 131K max output; 230B MoE / 10B active

Strengths

・Opus-class coding at roughly 1/25 the cost: 78% SWE-bench Verified (within 1-2 pts of Claude Sonnet 4) at $0.30/$1.20 per million tokens.
・Open weights (230B MoE, 10B active per token) for self-hosting; runs on SGLang, vLLM, Transformers, and NVIDIA NIM.
・Self-evolving training (100+ autonomous optimization cycles) and strong agentic tool use (GDPval-AA Elo 1495), handling 30-50% of MiniMax's own RL research workflow.

Weaknesses

・License is 'faux open-source': Modified-MIT requires written MiniMax authorization for commercial use, drawing community pushback.
・Text-only - no native vision or audio; image understanding must be proxied through a function call to another model.
・Graduate-level reasoning (GPQA Diamond) is well below its coding scores, and ~35-46 tokens/s throughput is slow for a 10B-active model.

Competitor Comparison

Model	Arena	SWE	GPQA	Price
Claude Sonnet 4 (Anthropic)	N/A	SWE-bench Verified ~79%	N/A	$3/$15
MiniMax M2.5 (MiniMax)	N/A	SWE-bench Verified 80.2%	N/A	$0.30/$1.20
GPT-5.3-Codex (OpenAI)	N/A	SWE-Pro ~56%	N/A	N/A

Overview

MiniMax M2.7 is an open-weight coding and agentic model from Shanghai-based MiniMax, released March 18, 2026 (weights April 12, 2026). It is a 230-billion-parameter Mixture-of-Experts model that activates only 10 billion parameters per token, scoring 78% on SWE-bench Verified and 56.22% on SWE-Pro - matching or beating GPT-5.3-Codex on the harder coding evals at roughly 1/25 the cost of Claude Opus 4.6. MiniMax positions it as a self-evolving model that ran 100+ autonomous optimization cycles during training, and as a production-grade coding agent for Cline, OpenCode, and Kilo harnesses. Its main caveats are a non-commercial Modified-MIT license and text-only modality.

Benchmarks & Performance

Coding and agentic benchmarks (MiniMax self-reported, cross-checked by Artificial Analysis and OpenRouter): SWE-bench Verified 78%, SWE-Pro 56.22%, SWE Multilingual 76.5%, Multi-SWE-Bench 52.7%, Terminal-Bench 2 57.0%, VIBE-Pro 55.6%, MLE-Bench Lite 66.6% (tying Gemini 3.1 Pro), GDPval-AA Elo 1495. Artificial Analysis gives an Intelligence Index of 50. Note that on SWE-bench Verified, M2.7 (78%) actually scores slightly below M2.5 (80.2%) - MiniMax frames this as a shift toward agentic and multilingual optimization. Graduate-level reasoning (GPQA Diamond) is notably lower, reflecting the code/agent training focus.

Detailed Comparison

**MiniMax M2.7 vs Claude Sonnet 4 (Anthropic).** M2.7 matches Sonnet 4 on SWE-bench Verified (78% vs ~79%) at one-tenth the price ($0.30/$1.20 vs $3/$15), but Sonnet 4 is multimodal and stronger on general reasoning. **MiniMax M2.7 vs MiniMax M2.5.** M2.7 adds self-evolution and agentic gains (Terminal-Bench 2 57% vs 47.9%) but slips slightly on SWE-bench Verified (78% vs 80.2%); pricing is identical. **MiniMax M2.7 vs GPT-5.3-Codex (OpenAI).** The two land within a point on SWE-Pro (~56%), but M2.7 is open-weight and far cheaper, while Codex integrates more tightly with OpenAI's coding stack.

Community Feedback

Developers highlight M2.7's Opus-class coding at a fraction of the cost and its open weights for self-hosting. The biggest controversy is the license: the 'Modified-MIT' tag with a required commercial authorization was widely criticized as 'faux open-source.' Reviewers also note slow throughput (~35-46 tok/s) and that graduate-level reasoning lags its coding strength. Independent replication on the harder evals remains limited.

Use Cases

1. **Autonomous code repair and refactoring.** 78% SWE-bench Verified means M2.7 can read a codebase, locate a bug, and produce a working fix for roughly four of five real GitHub issues - ideal for self-hosted coding agents. 2. **Multi-step agentic workflows.** Strong tool use (GDPval-AA Elo 1495) and Terminal-Bench 2 (57%) suit live debugging, root-cause analysis, and end-to-end document generation. 3. **Cost-sensitive production coding.** At $0.30/$1.20 per million, it undercuts every closed competitor at the same benchmark tier, making it attractive for high-volume CI and internal developer tools. 4. **When NOT to use.** Multimodal tasks (no native vision/audio - use MiniMax's M3 line), graduate-level reasoning, and latency-sensitive interactive chat (slow throughput).

Latest News

- **Released March 18, 2026:** MiniMax M2.7, 230B MoE / 10B active, open weights. - **Weights April 12, 2026:** published on Hugging Face under Modified-MIT (commercial use needs written authorization). - **Self-evolution:** ran 100+ autonomous optimization cycles during training, a 30% internal benchmark gain. - **Pricing:** $0.30/$1.20 per million (cached input $0.06); high-speed variant ~$0.60/$2.40.

Sources

Analysis generated: 2026-08-24