The brand new MiniMax M2.7 matches prime closed fashions in benchmarks.
The mannequin underwent a fairly license shift to limit business use post-release.
The transfer was justified as a strategy to stop service suppliers from “nerfing” the mannequin when providing it to builders.
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.
Picture: Minimax
It is a 230B-parameter Combination of Consultants mannequin with solely 10B lively per inference cross, so that you get frontier-level output with out paying frontier-level compute. MiniMax mentioned it was the primary mannequin to take part in its personal improvement—an inside 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 group misplaced it
However shortly after the weights dropped, the Chinese language AI lab MiniMax quietly up to date the phrases: business use now requires written authorization from MiniMax.
Non-commercial use stays free and unrestricted. Analysis, private tasks, fine-tuning on your personal setup—none of that modified. However in case you’re working a hosted service or constructing a business 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 business use by definition. Calling it “Modified-MIT” whereas limiting business use is, to place it charitably, complicated.
Ryan Lee, MiniMax’s Head of Developer Relations, posted an in depth response fairly than the same old company non-answer. His clarification: bad-faith internet hosting suppliers had been deploying degraded variations of earlier MiniMax fashions—mistaken templates, aggressive quantization, generally 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 consumer will get a nasty 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 strategy to push again on any of that,” he added. “If the license has edge instances that harm authentic group use, inform us. We would fairly repair the textual content than defend it.”
This matches a wider sample. MiniMax constructed its developer fame on absolutely open releases—M2 beneath MIT in October 2025, M2.5 beneath 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 Trade in January 2026, elevating round $620M with Alibaba and Abu Dhabi’s sovereign wealth fund amongst its backers.
Different Chinese language corporations, which dominate the open-source area, are additionally testing the waters of close-sourcing AI. Alibaba’s Qwen workforce reportedly shifted towards proprietary improvement after senior management departures, based on the Monetary Instances, Xiaomi additionally launched its new MiMo v2 fashions beneath an in depth supply license. The shorthand that Chinese language labs are open and U.S. labs are closed now not holds.
For these desirous about utilizing it commercially, Lee says the authorization course of might be quick and cheap.
Day by day Debrief E-newsletter
Begin daily with the highest information tales proper now, plus authentic options, a podcast, movies and extra.
