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

MiniMax M2.7 is here, the weights are on Hugging Face, and it’s legally taking on the very best closed designs out today.

The numbers: 56.22% on SWE-Pro (a standard for software application engineering jobs), almost matching Claude Opus 4.6; 57.0% on Terminal Bench 2. An ELO of 1495 on GDPval-AA (a standard for real-world understanding work jobs throughout tasks). For context: This is the greatest amongst open-weight designs, just a little listed below Opus 4.6, Sonnet 4.6, and GPT-5.4.

It’s a 230B-parameter Mix of Specialists design with only 10B active per reasoning pass, so you get frontier-level output without paying frontier-level calculate. MiniMax stated it was the very first design to take part in its own advancement– an internal variation ran 100+ self-governing rounds of self-optimization, reworded its own scaffold, and came out 30% much better. No human in the loop.

Then the license altered, and the neighborhood lost it

However soon after the weights dropped, the Chinese AI laboratory MiniMax silently upgraded the terms: industrial usage now needs written permission from MiniMax.

Non-commercial usage remains totally free and unlimited. Research study, individual tasks, tweak for your own setup– none of that altered. However if you’re running a hosted service or developing a business item, you remain in “requirements permission” area now.

Hacker News and a Hugging Face conversation thread filled quickly with designers calling it out. The particular friction point is this: MiniMax is identifying the license “MIT-style,” however MIT allows industrial usage by meaning. Calling it “Modified-MIT” while limiting industrial usage is, to put it charitably, puzzling.

Ryan Lee, MiniMax’s Head of Designer Relations, published a comprehensive reaction instead of the typical business non-answer. His description: bad-faith hosting service providers had actually been releasing abject variations of previous MiniMax designs– incorrect design templates, aggressive quantization, often not even MiniMax’s real design– then letting users leave believing MiniMax ships average work.

” They leave believing MiniMax is mid,” Lee composed. “We get the reputational expense, the user gets a disappointment, and the major hosting service providers who do the work appropriately get hushed in the sound.”

” A totally liberal license suggested we had no other way to press back on any of that,” he included. “If the license has edge cases that harm genuine neighborhood usage, inform us. We ‘d rather repair the text than protect it.”

This fits a broader pattern. MiniMax developed its designer track record on completely open releases– M2 under MIT in October 2025, M2.5 under the exact same terms in February 2026. M2.7 is the very first break from that streak, and it came simply months after the business noted on the Hong Kong Stock Market in January 2026, raising around $620M with Alibaba and Abu Dhabi’s sovereign wealth fund amongst its backers.

Other Chinese business, which control the open-source area, are likewise checking the waters of close-sourcing AI. Alibaba’s Qwen group apparently moved towards exclusive advancement after senior management departures, according to the Financial Times, Xiaomi likewise launched its brand-new MiMo v2 designs under a close source license. The shorthand that Chinese laboratories are open and U.S. laboratories are closed no longer holds.

For those thinking about utilizing it commercially, Lee states the permission procedure will be quick and sensible.