The numbers: 56.22% on SWE-Pro (a benchmark for software engineering tasks), nearly matching Claude Opus 4.6; 57.0% on Terminal Bench 2. An ELO of 1495 on GDPval-AA (a benchmark for real-world knowledge work tasks across jobs). For context: This is the highest among open-weight models, only slightly below Opus 4.6, Sonnet 4.6, and GPT-5.4.
Image: MinimaxIt's a 230B-parameter Mixture of Experts model with only 10B active per inference pass, so you get frontier-level output without paying frontier-level compute. MiniMax said it was 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.
Then the license changed, and the community lost itNon-commercial use stays free and unrestricted. Research, personal projects, fine-tuning for your own setup—none of that changed. But if you're running a hosted service or building a commercial product, you're in “needs authorization” territory now.
Ryan Lee, MiniMax's Head of Developer Relations, posted a detailed response rather than the usual 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."
"A fully permissive license meant we had no way to push back on any of that," he added. "If the license has edge cases that hurt legitimate community use, tell us. We'd rather fix the text than defend it."
For those interested in using it commercially, Lee says the authorization process will be fast and reasonable.
