MiniMax · 2026-04-12 · major
MiniMax M2.7 — open-weights self-evolving agent model
MiniMax drops open weights for M2.7, a 229B/10B-active MoE agentic model previously API-only. Hits 56.22% on SWE-Bench Pro and 57.0% on Terminal Bench 2, with downloads already in the tens of thousands.
MiniMax's frontier agentic model is now downloadable — a 229B MoE that scores like GPT-5.3-Codex on SWE-Bench Pro.
Key specs
| License | Modified MIT |
|---|---|
| Active params | 10B |
| Context window | 200K tokens |
| Total parameters | 229B |
| Experts | 256 |
| Swe bench pro | 56.22% |
| Terminal bench 2 | 57.0% |
What is it?
M2.7 is MiniMax's flagship coding and agent model, now released as open weights on Hugging Face under a modified-MIT license. It is a Mixture-of-Experts text model with 229B total parameters, 10B active per token, 256 experts, and a 200K context window. The weights drop comes about three weeks after the original API-only launch on March 18.
How does it work?
Architecturally M2.7 is a sparse MoE built around three capability areas MiniMax calls software engineering, professional office work, and Agent Teams (native multi-agent collaboration). The 'self-evolving' framing refers to the training process: an internal version of M2.7 was given access to its own coding scaffold and ran 100+ rounds of analyze-failure → modify-code → re-evaluate, achieving a 30% performance lift on its own programming harness. The released weights ship in F32, BF16, and F8_E4M3 tensor formats and are supported in vLLM, SGLang, and Transformers out of the box.
Why does it matter?
Open-weights agent models in this size class are still rare — most labs keep their best agentic models behind APIs. M2.7's published benchmarks (56.22% on SWE-Bench Pro, 57.0% on Terminal Bench 2) put it within striking distance of GPT-5.3-Codex on the coding axis, and the modified-MIT license makes it deployable in commercial products. For teams that want frontier-tier agent quality without paying per-token, this is one of the highest-leverage downloads of the month.
Who is it for?
Teams building coding and agent products that want open-weights inference at frontier-class quality.
Try it
huggingface.co/MiniMaxAI/MiniMax-M2.7