MiniMax-M3 is MiniMaxAI's first vision-language model, built as a 428-billion-parameter mixture-of-experts architecture that activates 22 billion parameters per inference. Its language backbone inherits from MiniMax-M2.7 and spans 60 layers with 128 experts, paired with a proprietary block-sparse attention mechanism called MiniMax Sparse Attention (MSA) that aims to keep long-context reasoning computationally tractable. A CLIP-style vision transformer with 32 layers handles image inputs at resolutions ranging from 336×336 up to 2016×2016, enabling both image-to-text and video-to-text workflows on the same checkpoint. The model is trained and optimized on Hopper-generation GPUs and supports BF16 along with MXFP8 precision, giving downstream deployments flexibility around throughput and memory trade-offs.

With the cataloged API limit context window documented at the framework level and positioning that targets workloads such as long-form video understanding, multi-hour coding sessions, and design-heavy pipelines, MiniMax-M3 fits use cases where a single model has to ingest large documents, codebases, or extended video alongside instructions. The sparse attention design is the key enabler here, since it lets the model hold much more state in context without paying full quadratic attention cost, which is what makes hour-long video analysis and agentic coding loops practical rather than aspirational. MiniMax's own model catalog frames M3 as a frontier multimodal coding model, reinforcing that the intended audience is teams building coding agents, document or video analysts, and design tooling rather than lightweight chat assistants. The Hugging Face release under the MiniMaxAI organization signals a research-friendly distribution that integrators can load alongside mainstream serving stacks.