MiniMax releases M2.7, limits commercial use

MiniMax posted weights for its M2.7 model on Hugging Face and then revised the license to require written authorization for commercial use. Research, personal projects and fine-tuning for private deployments remain permitted under the revised terms.

M2.7 is a 230-billion-parameter mixture-of-experts model that activates about 10 billion parameters per inference pass. MiniMax published benchmark results showing 56.22% on SWE-Pro, 57.0% on Terminal Bench 2 and an ELO of 1495 on GDPval-AA, placing the release among the top open-weight performers and near several leading closed models.

MiniMax said an internal version of M2.7 ran more than 100 autonomous rounds of self-optimization and rewrote its own scaffold, improving performance by roughly 30% without human intervention.

Shortly after the weights appeared on Hugging Face, the lab updated the license so commercial use requires written authorization. Noncommercial use remains free and unrestricted. Developers in Hugging Face discussion threads and other developer forums questioned MiniMax’s description of the license as “MIT-style,” noting that the standard MIT license permits commercial use.

Ryan Lee, MiniMax’s head of developer relations, wrote that some hosting providers had been deploying degraded or altered versions of previous MiniMax releases and presenting those services as the lab’s work. “They walk away thinking MiniMax is mid,” he wrote, adding that the lab had been left with the reputational cost while legitimate hosting providers struggled to compete. Lee said the license change provides a way to push back on bad-faith actors and invited feedback on edge cases that might harm legitimate community use.

MiniMax noted the authorization process for commercial users will be fast and reasonable. The company released earlier models M2 and M2.5 under permissive MIT terms in October 2025 and February 2026; M2.7 is the first release under the new licensing approach. MiniMax listed on the Hong Kong Stock Exchange in January 2026 and raised about $620 million, with investments that include Alibaba and Abu Dhabi’s sovereign wealth fund.

Several Chinese AI labs have moved toward more restrictive distribution for newer models. Developers and hosting providers have said controlled releases can protect model quality and brand reputation, while other developers and providers have said such restrictions reduce interoperability and create uncertainty for startups building commercial services on open weights.

Developers pressed MiniMax for clearer license text and a faster authorization process. MiniMax reiterated that noncommercial activities remain free and said it will consider revising license language if it creates unintended barriers for legitimate research or hobbyist use.