M3 is a cheap multimodal model with a 1M window, priced from $0.30/$1.20 per 1M tokens. Its low hallucination rate comes from attempting only 31% of questions, so check recall needs before adopting it.
Key takeaways · AI-distilled
MiniMax-M3 is the company's first multimodalA model that works with more than text — reading images, audio, or video, and sometimes generating them too.Full definition → M-series model, adding image and video input and raising the context windowThe maximum amount of text a model can consider at once — its working memory for the current conversation or task.Full definition → from M2.7's 200K to 1M tokenThe chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.Full definition →, while scoring 55 versus 50 on the Intelligence Index.
Gains over M2.7 include HLE 28% to 37%, GPQA Diamond 87% to 93% and IFBench 76% to 83%, with a small SciCode regression from 47% to 45%.
On AA-Omniscience M3 attempts only 30.9% of questions, the lowest among peers, which produces both a low hallucinationWhen a model states something false with full confidence — inventing facts, citations, or APIs that don't exist.Full definition → rate (16.1%) and low accuracy (15.0%).
At writing, weights were not yet released; MiniMax said they would follow within about 10 days, and M2.7's weights had shipped under a commercially restricted license.
Terms in this piece · Glossary
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
hallucination — When a model states something false with full confidence — inventing facts, citations, or APIs that don't exist.
multimodal — A model that works with more than text — reading images, audio, or video, and sometimes generating them too.
token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
