MiniMax Open-Sources M3.1 Flash Model Weights with 428B Total Parameters
On October 23, MiniMax announced the open-sourcing of its MiniMax M3.1 Flash model weights. The model has 428B total parameters and 23B activated parameters, delivering comprehensive capability improvements over the previous M3 generation. Through full-model sparse attention, quantization-aware training, and power-aware scheduling optimizations, M3.1 Flash achieves higher inference efficiency.
Specific optimizations include replacing the first three full-attention layers with sparse attention, migrating routing MoE experts to W4A4 (excluding shared experts), shifting the sparse attention indexer and main branch from Q8KV8 to Q8KV4, and dynamically managing compute card power to sustain high utilization and hardware efficiency.
M3.1 Flash went live on October 19 and processed over 100 trillion tokens from real users within 72 hours. MiniMax is working with domestic and international chip vendors including Huawei Ascend, T-Head, Kunlunxin, Moore Threads, Biren, Hygon, Muxi, MetaX, NVIDIA, and AMD, as well as inference and cloud platforms such as Together AI, Fireworks, Nebius, Baseten, Modal, and DigitalOcean to enable day-one integration.
In addition, frameworks and open-source communities including vLLM, SGLang, BAAI FlagOS, Hugging Face, and ModelScope will join MiniMax in promoting the model's deployment across different chips, cloud platforms, and development environments, further expanding the global AI open-source ecosystem.
