What happened
MiniMax published the M3.1-Flash card on October 17, 2026. The card says the model is a native multimodal mixture of experts, about 428 billion parameters in total, with 23 billion activated per token. Inputs are text, image, and video. Output is text. Context is up to 1 million tokens. The license field is other, and the license name is minimax-community. On MiniMax's table, Terminal-Bench 4.0 is 38.1 for this checkpoint, 42.3 for DeepSeek V4.1-Flash-1017, and 51.8 for Opus 5. DeepSWE v1.1 is 73.7, 77.0, and 74.0 on those same three columns. HLE without tools, on the text-only subset, is 43.4, 42.6, and 55.6. MiniMax says the Terminal-Bench run used Codex 0.149.1 and the DeepSWE run used mini-SWE-agent 2.4.5.
Why it matters
This is another coding and agent checkpoint with 23 billion parameters active, the same active count Beam published the day before under Apache 2.0. The license here is MiniMax Community, which is not Apache. On the one table MiniMax printed, DeepSeek's October 17 Flash checkpoint is ahead on Terminal-Bench 4.0 and on DeepSWE v1.1. Those columns are MiniMax's measurements of the other models, not a second lab's rerun. A team choosing a file to serve can compare the license and the active size. A team choosing an API still has to read DeepSeek's own table, which does not include this checkpoint.
The card says three things changed from M3. Sparse attention replaces the first three full-attention layers. The routed experts and the attention paths were quantized during training. A DSpark draft head replaces an earlier speculative-decoding setup, and the draft head is in the release.
Serving notes on the card name SGLang, vLLM, and TokenSpeed. The card says Transformers support was still in progress when the page was written.