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

MiniMax M2.7 is right here, the weights are on Hugging Face, and it is legitimately competing with the most effective closed fashions out proper now.

The numbers: 56.22% on SWE-Professional (a benchmark for software program engineering duties), almost matching Claude Opus 4.6; 57.0% on Terminal Bench 2. An ELO of 1495 on GDPval-AA (a benchmark for real-world data work duties throughout jobs). For context: That is the best amongst open-weight fashions, solely barely under Opus 4.6, Sonnet 4.6, and GPT-5.4.

It is a 230B-parameter Combination of Specialists mannequin with solely 10B lively per inference go, so that you get frontier-level output with out paying frontier-level compute. MiniMax stated it was the primary mannequin to take part in its personal improvement—an inner model ran 100+ autonomous rounds of self-optimization, rewrote its personal scaffold, and got here out 30% higher. No human within the loop.

Then the license modified, and the neighborhood misplaced it

However shortly after the weights dropped, the Chinese language AI lab MiniMax quietly up to date the phrases: industrial use now requires written authorization from MiniMax.

Non-commercial use stays free and unrestricted. Analysis, private initiatives, fine-tuning in your personal setup—none of that modified. However should you’re operating a hosted service or constructing a industrial product, you are in “wants authorization” territory now.

Hacker Information and a Hugging Face dialogue thread stuffed up quick with builders calling it out. The particular friction level is that this: MiniMax is labeling the license “MIT-style,” however MIT permits industrial use by definition. Calling it “Modified-MIT” whereas limiting industrial use is, to place it charitably, complicated.

Ryan Lee, MiniMax’s Head of Developer Relations, posted an in depth response moderately than the standard company non-answer. His rationalization: bad-faith internet hosting suppliers had been deploying degraded variations of earlier MiniMax fashions—fallacious templates, aggressive quantization, typically not even MiniMax’s precise mannequin—then letting customers stroll away pondering MiniMax ships mediocre work.



“They stroll away pondering MiniMax is mid,” Lee wrote. “We get the reputational invoice, the person will get a foul expertise, and the intense internet hosting suppliers who do the work correctly get drowned out within the noise.”

“A totally permissive license meant we had no method to push again on any of that,” he added. “If the license has edge instances that harm respectable neighborhood use, inform us. We would moderately repair the textual content than defend it.”

This suits a wider sample. MiniMax constructed its developer fame on totally open releases—M2 underneath MIT in October 2025, M2.5 underneath the identical phrases in February 2026. M2.7 is the primary break from that streak, and it got here simply months after the corporate listed on the Hong Kong Inventory Change in January 2026, elevating round $620M with Alibaba and Abu Dhabi’s sovereign wealth fund amongst its backers.

Different Chinese language firms, which dominate the open-source area, are additionally testing the waters of close-sourcing AI. Alibaba’s Qwen group reportedly shifted towards proprietary improvement after senior management departures, based on the Monetary Occasions, Xiaomi additionally launched its new MiMo v2 fashions underneath an in depth supply license. The shorthand that Chinese language labs are open and U.S. labs are closed not holds.

For these serious about utilizing it commercially, Lee says the authorization course of will probably be quick and affordable.