Shanghai-based multimodal foundation model lab competing with OpenAI and Anthropic in text, video, speech, and music generation. Its open-weight M1 reasoning model uses a hybrid MoE plus lightning attention architecture, trained with a custom RL algorithm (CISPO), supporting up to 1M token context while outperforming DeepSeek-R1 on SWE-bench and FullStackBench.
MiniMax open-sources M3-MXFP8, a 428B-parameter multimodal MoE with a 1M-token context window, new sparse attention, and frontier coding scores — all runnable locally.