• Sources: primary, discussion
  • Thread: MiniMax M3.1 Flash open weights
  • Summary: The MiniMaxAI/MiniMax-M3.1-Flash repository lists a native multimodal mixture-of-experts model with about 428B total and 23B active parameters, text, image and video input, and a context window up to 1M tokens. It uses MiniMax Sparse Attention throughout, quantization-aware training, and a DSpark speculative decoding head, ships in mxfp8, and is licensed under minimax-community. Benchmark figures such as 73.7 percent on DeepSWE v1.1 are vendor-reported.
  • Why it matters: A large open-weight agentic coding model is now downloadable for vLLM and SGLang serving, though the license is custom rather than OSI.
  • Follow-up: Independent evaluations against the vendor benchmark table.