IT之家 · 智能时代 · 10/23/2026, 09:08:08
MiniMax Open-Sources M3.1 Flash Weights with 428B Total and 23B Active Parameters
MiniMax has open-sourced the weights for its M3.1 Flash model, featuring a sparse architecture with 428B total and 23B active parameters. The release enhances inference efficiency through full-model sparse attention, quantization-aware training, and power-aware scheduling, with day-one support from major chip vendors and cloud platforms.
SOURCE COVERAGEFull Original Report
On October 23, MiniMax announced the open-sourcing of model weights for MiniMax M3.1 Flash, featuring a total of 428B parameters with 23B active parameters.
Compared to the previous generation M3 model, M3.1 Flash delivers comprehensive improvements in model capabilities. Additionally, through optimizations such as full-model sparse attention, quantization-aware training (QAT), and power-aware scheduling, M3.1 Flash further enhances inference efficiency:
- Full-model sparse attention: The first three full-attention layers are replaced with sparse attention;
- QAT: Routed MoE experts are migrated to W4A4 (excluding shared experts), while the sparse attention indexer and main branch are migrated from Q8KV8 to Q8KV4;
- Power-aware scheduling: By precisely allocating power consumption across compute cards, this approach keeps compute cards consistently operating at high capacity, maximizing hardware efficiency.
M3.1 Flash launched on October 19, processing over 100 trillion tokens from real users within 72 hours.
MiniMax announced partnerships with domestic and international chip manufacturers—including Huawei Ascend, T-Head, Kunlunxin, Moore Threads, Biren Technology, Hygon, MetaX, Iluvatar CoreX, NVIDIA, and AMD—as well as inference and cloud service platforms such as Together AI, Fireworks, Nebius, Baseten, Modal, and DigitalOcean, to jointly facilitate day-one integration support for the open-source release of M3.1 Flash.
Furthermore, inference frameworks and open-source communities including vLLM, SGLang, FlagOS by Beijing Academy of Artificial Intelligence (BAAI), Hugging Face, and ModelScope will collaborate with MiniMax to promote the application of the model across various chips, cloud platforms, and development environments, continuously expanding the global AI open-source ecosystem.
IT Home provides the following open-source addresses: