Chinese AI laboratory MiniMax has released a new open-weight agent model that rivals leading closed systems on key benchmarks, then updated its licence to require written authorisation for commercial use, triggering a sharp reaction from the developer community.
MiniMax M2.7, a 230-billion-parameter mixture-of-experts model, scored 56.22% on SWE-Pro, a software engineering benchmark, placing it close to Claude Opus 4.6, and 57.0% on Terminal Bench 2, a measure of autonomous task completion in command-line environments.
Mixture-of-experts is an architecture that activates only a subset of a model's parameters during each inference pass; M2.7 uses 10 billion of its 230 billion parameters per pass, delivering high-end output at lower computational cost.
The company said an internal version of the model completed more than 100 autonomous rounds of self-optimisation, rewrote its own scaffolding, and improved performance by 30% without human involvement.
Shortly after the weights were published on AI model repository Hugging Face, MiniMax updated the licence terms to restrict commercial deployment without prior approval, while leaving non-commercial research, personal use and fine-tuning unrestricted.
Developers on Hacker News and Hugging Face forums objected quickly, noting that MiniMax was describing the licence as MIT-style despite the MIT licence permitting commercial use by definition.
Ryan Lee, MiniMax's head of developer relations, said the change was prompted by hosting providers deploying degraded versions of earlier MiniMax models using incorrect templates or aggressive compression, then presenting the results as representative of MiniMax's work.
"They walk away thinking MiniMax is mid," Lee wrote. "We get the reputational bill, the user gets a bad experience, and the serious hosting providers who do the work properly get drowned out in the noise."
Lee said the authorisation process for commercial users would be fast and reasonable, and invited feedback on licence language that might affect legitimate community use.
M2.7 marks the first departure from fully open licensing in MiniMax's recent model releases: its M2 and M2.5 models, released in October 2025 and February 2026 respectively, both shipped under standard MIT terms.
Related reading
- Memento-Skills lets AI agents rewrite their own capabilities without retraining underlying models
- Pixel Societies: the hackathon project building pixel-art AI twins to test human compatibility
- Microsoft confirms 365 Copilot agent with OpenClaw-like local capabilities ahead of Build showcase
The shift follows MiniMax's listing on the Hong Kong Stock Exchange in January 2026, which raised approximately $620 million with Alibaba and Abu Dhabi's sovereign wealth fund among its backers.
The licence change reflects a wider trend among Chinese AI laboratories, which have previously been characterised as more open than their US counterparts. Alibaba's Qwen team has reportedly moved toward proprietary development following senior departures, and Xiaomi released its MiMo v2 models under a closed licence.
The recap
- MiniMax released a state-of-the-art artificial intelligence agent model.
- The company updated the model's licence after the initial release.
- Company described the change in an announcement; no timeline provided.
