MiniMax released M3.1 Flash on October 19, 2026, a 428-billion-parameter Mixture-of-Experts (MoE) model that activates only 23 billion parameters per token, achieving over 200 tokens per second and costing $0.15 per million input tokens — a 70% reduction from its predecessor. The model leverages sparse attention through MiniMax Sparse Attention (MSA), replacing the first three full-attention layers with sparse alternatives to enable a practical 1-million-token context window. For faster generation, it uses DSpark speculative decoding, a new scheme with a dedicated draft head that improves throughput by reducing token rejection rates, reporting median speeds of 170+ tokens per second on NVIDIA Blackwell hardware (B200, B300, GB200, GB300). Quantization-aware training (QAT) applies NVFP4 (W4A4) precision to MoE experts and Q8KV4 to the sparse-attention indexer and main branch, preserving near-full-precision accuracy while minimizing memory usage. The model supports five tunable reasoning levels — low, medium, high, xhigh, and max — allowing developers to balance latency and reasoning depth per task. Benchmarks show strong performance in agentic applications: 82% on CyberGym, 73.7% on DeepSWE v1.1, and 61.7% on OSWorld V2. Weighing in at 29.6 on Agents' Last Exam, the model still trails on highly complex multi-step tasks. It runs on vLLM, SGLang, and TokenSpeed, and is available on Hugging Face with API access priced at $0.15 and $0.60 per million input and output tokens, respectively.