MiniMax M2.7 review: a self-improving open model for agents

MiniMax M2.7 (minimax-m2.7) is an open-weight model built for coding and multi-agent automation for much less than the top closed models cost to run. Agent-focused training plus an efficient MoE design put it close to leading closed models on agent benchmarks and in independent tests. Two caveats: there is little public evidence on how it does outside English and Chinese, and the weights come with MiniMax's own license.

MiniMax has since released MiniMax M3, which adds image and video input. M2.7 is still served and remains the text-only option in the family.

What it is

MiniMax, the Shanghai AI company, released M2.7 as open weights on March 18, 2026.

• Architecture: Mixture-of-Experts, 230B total parameters, about 10B active per token (chosen from 256 experts)
• Context: about 200K tokens
• Tool and function calling supported
• Input and output: text

The headline feature is what MiniMax calls self-evolution. During training, the model ran more than 100 rounds of optimizing its own agent scaffold, with no human in the loop, and MiniMax reports a 30% gain on its internal evaluations from that alone. It is the first MiniMax model to take an active part in its own training.

Strengths

• Agent and coding benchmarks: SWE-Pro 56.22%, Terminal Bench 2 57.0%, SWE-bench Verified 78%. Its Artificial Analysis Intelligence Index is 50, against a median of 31 for open-weight models.
• Cost efficiency: in an independent test by Kilo, it kept about 90% of Claude Opus 4.6's quality at much lower cost.
• Built for multi-agent work: MiniMax trained for role boundaries, adversarial reasoning and protocol adherence. On its internal MM Claw evaluation it followed 97% of 40 complex skills, with 62.7% overall accuracy.
• Usable speed: Artificial Analysis measured 46.5 tokens per second and 2.28 seconds to first token on MiniMax's own API.
• Open weights: you can self-host it or fine-tune it.

Weak spots

• License: the weights are under MiniMax's own license. Read it before you self-host for a commercial product. Calls through FastMetal's API fall under FastMetal's terms.
• Languages beyond English and Chinese: we found no public benchmark for, say, Japanese. Test your own prompts.
• General coding: BenchLM reports a coding score of 53.7 for M2.7 against 77.2 for Claude Sonnet 4.5. The agentic strength does not carry over to everyday code completion.
• Verbosity: several reviews call it a bit wordy. Asking for an explicit output format helps.
• Vibe coding: one report has its vibe-coding subscore below its predecessor M2.5's.

How it compares

Each figure comes from a different source with different conditions. Read it as a general direction.

Cost

M2.7 is billed per million input and output tokens, and it is one of the cheaper open-weight models in the catalog. That makes it a good fit for agent loops that run many calls, and for proofs of concept where you want to check quality without a large bill. The current rate is on the model page.

Outside English

Neither MiniMax nor third parties have published a benchmark for languages like Japanese. Training data is presumably mostly English and Chinese, so fluency and vocabulary in other languages may vary. Structured output (code, tool calls) depends less on language and tends to hold up. Before you ship prose in another language, test it on your real domain and audience.

Where it fits

Good fit: multi-agent automation, coding assistance and debugging, test generation, long multi-step tool calling, low-cost proofs of concept and research, and SWE-style benchmark comparisons.

Poor fit: polished prose in languages other than English and Chinese (test first), commercial self-hosting without checking the license, and general chat or creative writing where you would otherwise use a frontier model.

FAQ

Can I use MiniMax M2.7 in a commercial project?

Through FastMetal's API, FastMetal's terms apply. If you self-host the weights commercially, check MiniMax's license first.

How is M2.7 different from MiniMax-Text-01 (M1)?

Text-01 is a 456B general model with up to 4M tokens of context. M2.7 is smaller (230B, 10B active) and trained specifically for agents and coding. FastMetal serves M2.7 but not Text-01.

When does the 200K context help?

Loading a large codebase, agents that work across many documents, long log analysis: anything that packs a lot into one request. Speed and quality can shift as context grows, so test with real data.

Try it on FastMetal

curl https://api.fastmetal.ai/v1/chat/completions \
  -H "Authorization: Bearer $FASTMETAL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "minimax-m2.7",
    "messages": [{"role": "user", "content": "Write a pytest test for a function that parses ISO 8601 dates."}]
  }'

The same key reaches minimax-m3 and every other model in the catalog.