MiniMax Drops State-of-the-Art AI Agent Model—Then Quietly Changes the License

MiniMax M2.7 is here, the weights are connected Hugging Face, and it's legitimately competing with the champion closed models retired close now.

The numbers: 56.22% connected SWE-Pro (a benchmark for bundle engineering tasks), astir matching Claude Opus 4.6; 57.0% connected Terminal Bench 2. An ELO of 1495 connected GDPval-AA (a benchmark for real-world cognition enactment tasks crossed jobs). For context: This is the highest among open-weight models, lone somewhat beneath Opus 4.6, Sonnet 4.6, and GPT-5.4.

Image: Minimax

It's a 230B-parameter Mixture of Experts exemplary with lone 10B progressive per inference pass, truthful you get frontier-level output without paying frontier-level compute. MiniMax said it was the archetypal exemplary to enactment successful its ain development—an interior mentation ran 100+ autonomous rounds of self-optimization, rewrote its ain scaffold, and came retired 30% better. No quality successful the loop.

Then the licence changed, and the assemblage mislaid it

But soon aft the weights dropped, the Chinese AI laboratory MiniMax quietly updated the terms: commercialized usage present requires written authorization from MiniMax.

Non-commercial usage stays escaped and unrestricted. Research, idiosyncratic projects, fine-tuning for your ain setup—none of that changed. But if you're moving a hosted work oregon gathering a commercialized product, you're successful “needs authorization” territory now.

Hacker News and a Hugging Face treatment thread filled up accelerated with developers calling it out. The circumstantial friction constituent is this: MiniMax is labeling the licence "MIT-style," but MIT permits commercialized usage by definition. Calling it "Modified-MIT" portion restricting commercialized usage is, to enactment it charitably, confusing.

Ryan Lee, MiniMax's Head of Developer Relations, posted a elaborate effect alternatively than the accustomed firm non-answer. His explanation: bad-faith hosting providers had been deploying degraded versions of erstwhile MiniMax models—wrong templates, assertive quantization, sometimes not adjacent MiniMax's existent model—then letting users locomotion distant reasoning MiniMax ships mediocre work.

"They locomotion distant reasoning MiniMax is mid," Lee wrote. "We get the reputational bill, the idiosyncratic gets a atrocious experience, and the superior hosting providers who bash the enactment decently get drowned retired successful the noise."

"A afloat permissive licence meant we had nary mode to propulsion backmost connected immoderate of that," helium added. "If the licence has borderline cases that wounded morganatic assemblage use, archer us. We'd alternatively hole the substance than support it."

This fits a wider pattern. MiniMax built its developer reputation connected afloat unfastened releases—M2 nether MIT successful October 2025, M2.5 nether the aforesaid presumption successful February 2026. M2.7 is the archetypal interruption from that streak, and it came conscionable months aft the institution listed connected the Hong Kong Stock Exchange successful January 2026, raising astir $620M with Alibaba and Abu Dhabi's sovereign wealthiness money among its backers.

Other Chinese companies, which predominate the open-source space, are besides investigating the waters of close-sourcing AI. Alibaba's Qwen squad reportedly shifted toward proprietary improvement aft elder enactment departures, according to the Financial Times, Xiaomi besides released its caller MiMo v2 models under a adjacent root license. The shorthand that Chinese labs are unfastened and U.S. labs are closed nary longer holds.

For those funny successful utilizing it commercially, Lee says the authorization process volition beryllium accelerated and reasonable.