Minimax M3

MiniMax-M3 scores 55 on Artificial Analysis's Intelligence Index, just ahead of the open weights models Kimi K2.6 and MiMo-V2.5-Pro at 54 each (artificialanalysis.ai). The headline's condition is about status rather than performance. When the article was published, M3 was served through MiniMax's own API, SiliconFlow, GMI and Novita, but its weights had not been released. MiniMax said they would follow within about 10 days, and Artificial Analysis noted that M2.7's weights had come under a commercially restricted license. A MiniMax-M3 repository is now public on Hugging Face under a MiniMax Community License. It grants free use for non-commercial purposes, while commercial use requires a Built with MiniMax M3 notice and, for products with more than $20 million in yearly revenue, prior written authorization from MiniMax.

M3 is the first multimodal model in MiniMax's M series, adding image and video input to the text-only M2.7, and its context window grows from 200,000 to 1 million tokens. It improves on M2.7 on most evaluations, including Humanity's Last Exam from 28% to 37%, GPQA Diamond from 87% to 93% and IFBench from 76% to 83%, with a small drop on SciCode from 47% to 45%. It used about 91 million output tokens to run the index against M2.7's 87 million, so its 5-point gain over M2.7's index score of 50 did not come from much longer answers.

On GDPval-AA, a benchmark of real-world tasks across 44 occupations, M3 scores about 1670, behind Claude Opus 4.8 (max) at 1890 and GPT-5.5 (xhigh) at 1769, and level with Claude Sonnet 4.6 at 1676. AA-Omniscience shows a distinctive habit. M3 attempted only 30.9% of questions, the fewest among its peers. Declining to answer keeps its hallucination rate low at 16.1%, but it also holds accuracy down at 15.0%. Pricing is $0.30 per million input tokens and $1.20 per million output tokens up to 512,000 tokens of context, doubling to $0.60 and $2.40 beyond that.