I need to be ruthlessness-maxing I have been very soft on MiniMax, largely because many of their people follow me and seem to be good faith researchers. They're getting punished by the market anyway, so here goes. (inhales) What I saw, from the outside:
bad strategic vision, or rather absence of it. Smallness. I can tell what DS and Kimi, even StepFun are trying to achieve and it goes beyond a coding plan business, there's a scientific research program or even philosophical and aesthetic project behind them all. GLM is a slow-but-steady commercial program with good execution, long-term ambition and academic foundation. I have no clue what you were going for, it felt like merely chasing the latest monetizable trend or just a claim to revenue. Too much cope about perils of high-risk architecture research after M1 with "Lightning Attention" didn't do well; too cowardly and long-lasting a retreat into a long-derisked GQA design, empty triumphalism that you optimized and it's "very fast" (uh-huh, until KV cache memory pressure becomes the bottleneck, which it does in agentic workloads), even some sneering about NSA's scaling failure (of course DeepSeek just went back to the drawing board and invented DSA and then 2 other designs). Spins that M2 was "mini actually, we just were surprised how good it works so skipped a full-size one" and "still Opus-tier fr fr, check out SWA-bench" despite obvious overfit which became more obvious as harder and more OOD evaluations started coming out. Reliance on raw distillation – 100x more than DS as per Anthropic's report, and unlike them, not using Claude creatively to reverse engineer its reasoning or reward profile, but a raw 2024-coded attempt at capability theft. Presumably, underdeveloped culture of telemetry and internal evals, which is why you constantly make claims that real user experience doesn't match, like early-fusion multimodality that falls apart completely unlike Gemini's, or even Kimi's late fusion. I also suspect bad scaling laws. Flexing 996 in 2026? Over-the-top and yet amateurish self-promotion on social media that looks desperate. Your CEO had said he took inspiration from Wenfeng on openness, but it might have been a PR move rather than Yang Zhilin's "we drop our armor, and dare the world find a flaw" after all. Despite this, you have strong sides. I hoped this distillation will be metabolized into robust capabilities via your RL flywheel you are clearly proud of. You know AA benchmarks where you do unusually well, like Omniscience non-hallucination, or agentic ones that your RL does cover. M3 isn't a crappy model for its scale, and might have a nontrivial potential as a base. You dared to ship a simple and apparently efficient attention. You have raised capital that should have sufficed for a more significant acceleration. But overall… …NGMI energy. You need to change pace by M3.1. Reform yourselves like Kimi did all the way back then in January 2025, or get what you deserve. They will be getting what they deserve, too.
Yes I feel bad about it because clearly MiniMax M3 is a technological leap over M2, and might well be the best open weights (provisionally) on the market but sorry, this engagement mode is not what gets me going. This doesn't scream "frontier AGI lab"
