~428B total / ~23B active (MoE)
Up to 1M tokens (1,048,576-token checkpoint configuration)
minimax-community
Open-weight native multimodal Mixture-of-Experts model for coding, long-horizon agents, computer use, cybersecurity, and professional workflows, supporting text, image, and video input with text output. MiniMax-M3.1-Flash retains M3's approximately 428B-total / 23B-active scale while introducing MiniMax Sparse Attention throughout all 60 transformer layers, replacing the three full-attention layers retained by M3. The text backbone has a 6,144-dimensional hidden state, 64 query heads, four key/value heads, three initial dense layers, and 57 MoE layers containing 128 routed experts plus one shared expert, with four routed experts activated per token. Its published configuration supports 1,048,576 positions and a 200,064-token vocabulary. The release uses quantization-aware training and ships a mixed compressed checkpoint whose configuration applies dynamic MXFP8 quantization while packing routed MoE experts in NVFP4 format. The main checkpoint contains 48 safetensors shards representing approximately 250.1 GB of tensors, alongside an approximately 2.3 GB DSpark speculative-decoding checkpoint. DSpark replaces M3's earlier EAGLE-like multi-token-prediction approach and is included for compatible serving frameworks. The supplied chat template supports low, medium, high, xhigh, and max reasoning-effort levels; MiniMax recommends temperature 1.0 and top-p 0.95. Publisher-reported results include 62.2 on Automation-Bench v1.0.6, 76.2 on BankerToolBench, 61.7 on OSWorld V2, 29.6 on Agents' Last Exam, 38.1 on Terminal-Bench 4.0, 60.3 on LOCA-Bench 256K, 73.7 on DeepSWE v1.1, 61.2 on NL2Repo-Bench, 82.2 on CyberGym, 91.0 on BabyVision with tools, 43.4 on Humanity's Last Exam with tools, and 82.5 on CharXiv RQ. MiniMax says online inference can exceed 200 output tokens per second, with median speeds above 170 tokens per second, while hosted agentic-task costs are approximately 70% below M3. MiniMax identified the earlier anonymous Space Bunny preview as an early version of M3.1 Flash. The custom MiniMax Community License permits non-commercial use, requires attribution and notice for most commercial use, and requires prior authorization when the associated products or services exceed $20 million in annual revenue.