MiniMax Quietly Bans Commercial Use of Its M2.7 Agent Model — Developers Call It 'Faux Open Source'

MiniMax has changed the license on its M2.7 self-evolving agent model from the standard MIT license to a “Modified-MIT” license that prohibits commercial use without written authorization from the company. The switch happened quietly—no blog post, no announcement thread—and developers only noticed when they read the updated model card on Hugging Face.

M2.7 is not a minor release. It scores 56.22% on SWE-Pro and 57.0% on Terminal Bench 2, making it a competitive agentic coding model capable of autonomous software engineering tasks. Its defining feature is self-improvement: the model can refine its own outputs through iterative feedback loops, a capability that makes it particularly valuable for long-horizon coding agents. The previous two releases—M2 in October 2025 and M2.5 in February 2026—both shipped under standard MIT. That track record is exactly why the reversal landed so badly.

MiniMax IPO’d on the Hong Kong Stock Exchange in January 2026, raising approximately $620 million. M2.7’s license change reads as a post-IPO monetization move: the company built community and adoption on the back of permissive licensing, then tightened the terms once the model was established. The pattern is familiar—Redis, Terraform, Elasticsearch all did versions of this—but the AI model space moves faster and the community backlash is sharper.

The technical problem with “Modified-MIT” is that it is not MIT. The OSI definition of open source requires that the license allow commercial use without restriction. A license that requires written authorization for commercial use fails the OSI Open Source Definition at criterion 6 (“No Discrimination Against Fields of Endeavor”). Calling it “Modified-MIT” is legally confusing and, critics argue, deliberately so—it gives the model the appearance of openness while retaining meaningful commercial control.

For developers, the immediate question is what this means for projects already built on M2 or M2.5 under MIT. Those releases remain MIT-licensed, and nothing about the M2.7 change retroactively affects code already built on earlier versions. But anyone who upgraded their agent stack to M2.7 expecting MIT terms—and many did, given the lineage—now needs to audit their use for commercial implications.

The broader pattern is worth naming. Chinese AI labs have made permissive licensing a competitive differentiator against closed US models. DeepSeek’s MIT license on R1 drove massive adoption. MiniMax’s reversal signals that at least some labs see that strategy as a phase, not a permanent commitment. If Qwen, InternLM, or other Chinese open releases follow with similar restrictions post-monetization, the “open” Chinese AI wave will look very different in retrospect.

MiniMax has not published a public explanation for the change.