~428B total / ~23B active (MoE)
1M tokens
minimax-community
Official open-weight native multimodal MiniMax flagship release positioned for long-horizon agentic work across coding, cowork, text, image, and video understanding. The Hugging Face model card describes MiniMax-M3 as a million-context multimodal Mixture-of-Experts model with about 428B total parameters and about 23B activated parameters, supporting both thinking and non-thinking modes depending on whether the task prioritizes deeper reasoning or lower-latency completion. The accompanying MiniMax Sparse Attention paper presents M3 as a production model built around blockwise sparse attention co-designed for efficient million-token inference, with group-specific Top-k key-value block selection and a GPU execution path that turns sparsity into practical speedups. Release materials emphasize native multimodal training from the first step rather than late fusion, strong coding and cowork capability on long-horizon agentic benchmarks, and large long-context efficiency gains versus MiniMax M2, including reported 9x prefill and 15x decode speedups at 1M context on the public model card while the technical report cites 14.2x prefill and 7.6x decoding wall-clock speedups on H800 for the MSA kernel versus dense GQA baselines. The card also highlights deployment support across Transformers, vLLM, and SGLang, with recommended inference settings and downloadable open weights for local serving.