MiniMax M3.1 Flash

MiniMax M3.1 Flash builds on MiniMax M3 with the same natively multimodal mixture-of-experts architecture — about 428 billion total parameters with 23 billion activated per token, accepting text, image and video input and returning text across a context window of up to 1 million tokens. Three changes drive the gains: sparse attention replaces the first three full-attention layers so attention is sparse throughout, quantization-aware training shrinks the routed experts and attention paths for cheaper serving, and DSpark speculative decoding ships with a draft head for compatible runtimes. It scores 76.2 on BankerToolBench, 73.7% on DeepSWE v1.1, 61.7 on OSWorld 2.0, 62.8 on Automation-Bench v1.0.6, 82.2 on CyberGym and 43.4 on HLE without tools, making it a strong fit for coding agents, computer-use workflows and cybersecurity tasks where a million-token working set matters. Recommended sampling is temperature 1.0 and top_p 0.95, with reasoning effort selectable from low through max.