IT Home On October 23, MiniMax Technology announced the open-source model weights for MiniMax M3.1 Flash. The total parameters are 428B, and the activated parameters are 23B.
Compared to the previous generation M3 model, the model capabilities of M3.1 Flash have been significantly improved. Additionally, through optimizations such as full-model sparse attention, quantitative perception training, and Power-aware scheduling, M3.1 Flash has further enhanced inference efficiency:
Full model sparse attention: The first three full attention layers are replaced with sparse attention;
QAT: Route MoE expert migration to W4A4 (except shared experts), sparse attention indexer and main branch are migrated from Q8KV8 to Q8KV4;
Power-aware scheduling： By carefully allocating the power consumption of the compute card, the compute card maintains a high computing power level continuously, maximizing hardware efficiency.
M3.1 Flash launched on October 19, and real users processed more than 100 trillion Tokens within 72 hours.
MiniMax has announced a collaboration with chip manufacturers both at home and abroad, such as Huawei昇腾, Pingtouge, Kunlun芯, Moore Thread, Biren, Hailong, MuXi, Tiantian, NVIDIA, and AMD, as well as inference and cloud service platforms like Together AI, Fireworks, Nebius, Baseten, Modal, and DigitalOcean. They will jointly promote the first-day open access to M3.1 Flash.
In addition, inference frameworks such as vLLM, SGLang, Zhiyuan FlagOS, Hugging Face, and MagicLlama, along with open-source communities, will also work with MiniMax to promote the application of models across various chips, cloud platforms, and development environments, continuously expanding the global AI open-source ecosystem.
IT Home provides the open-source address as follows:
Hugging Face: https://huggingface.co/MiniMaxAI/MiniMax-M3.1-Flash
Magic Bridge Community: https://modelscope.cn/models/MiniMax/MiniMax-M3.1-Flash