Open Weights, Closed Wallet: MiniMax's 'MIT-Style' License Gets a Commercial Upgrade After $620M Hong Kong IPO

MiniMax M2.7 dropped this week, and the benchmarks are genuinely impressive. We're talking 56.22% on SWE-Pro, 57.0% on Terminal Bench 2, and an ELO of 1495 on GDPval-AA. For context: highest among open-weight models, sitting just below Opus 4.6, Sonnet 4.6, and GPT-5.4. Basically, it flexed so hard on the leaderboard that other models started questioning their life choices.

The model itself? A 230B-parameter Mixture of Experts architecture with only 10B active per inference. Frontier-level output without frontier-level compute bills. It's like hiring a entire consultancy firm but only paying the one intern who actually does the work.

But here's where it gets interesting.

MiniMax claims M2.7 is the first model to participate in its own development. An internal version ran 100+ autonomous rounds of self-optimization, rewrote its own scaffold, and came out 30% better. No human in the loop. The model basically completed its own character arc like an anime protagonist—except instead of friendship, it discovered self-improvement.

Then the license changed.

Shortly after the weights hit Hugging Face, MiniMax quietly updated the terms. Commercial use now requires written authorization. Non-commercial stays free and unrestricted—research, personal projects, fine-tuning, all good. The classic open-source bait-and-switch, but make it AI.

But if you're building a hosted service or commercial product, you're in "ask permission" territory.

The community noticed fast. Hacker News and the Hugging Face discussion thread filled up with developers pointing out the obvious: MiniMax is calling this "Modified-MIT," but MIT permits commercial use by definition. That's a fun linguistic trick. It's like selling "sugar-free candy" that just uses a different definition of sugar.

Ryan Lee, MiniMax's Head of Developer Relations, posted a more detailed response than your typical corporate non-answer. His explanation: bad-faith hosting providers had been deploying degraded versions of previous MiniMax models—wrong templates, aggressive quantization, sometimes not even MiniMax's actual model—then letting users walk away thinking MiniMax ships mediocre work.

"They walk away thinking MiniMax is mid," Lee wrote. "We get the reputational bill, the user gets a bad experience, and the serious hosting providers who do the work properly get drowned out in the noise."

He added: "A fully permissive license meant we had no way to push back on any of that." He also left the door open: "If the license has edge cases that hurt legitimate community use, tell us. We'd rather fix the text than defend it." Fair play to actually responding instead of the usual corporate ghosting.

This breaks a streak. MiniMax built developer goodwill on fully open releases—M2 under MIT in October 2025, M2.5 under the same terms in February 2026. M2.7 is the first departure from that pattern. The vibes have shifted.

The timing isn't subtle either. MiniMax listed on the Hong Kong Stock Exchange in January 2026, raising around $620M with Alibaba and Abu Dhabi's sovereign wealth fund backing. Post-IPO license tightening—coincidence, probably. Sure. And my aunt's dog just happens to have the same name as my boss's stock portfolio.

MiniMax isn't alone in this pivot. Alibaba's Qwen team reportedly shifted toward proprietary development after senior leadership departures. Xiaomi released its MiMo v2 models under a closed license. The shorthand that Chinese labs are open and U.S. labs are closed? That narrative is getting stale fast. The great open-source east vs proprietary west divide is dissolving faster than my savings in a bull run.

For those genuinely interested in commercial deployment, Lee says the authorization process will be "fast and reasonable." Whether the community trusts that promise is another question. We've seen "fast and reasonable" before. It usually means fill out this 47-page form and wait six to eight quarters.