On October 23, MiniMax officially open-sourced the MiniMax M3.1Flash model weights. The model has 428B total parameters and 23B activated parameters, with comprehensive improvements in capability over the previous generation M3 model. According to official disclosures, within 72 hours of launch, M3.1Flash had processed over 100 trillion tokens for real users.
To improve inference efficiency, M3.1Flash introduces multiple optimizations: replacing the first three full attention layers with sparse attention; through quantization-aware training (QAT), migrating the routed MoE experts (except shared experts) to W4A4, and adjusting the sparse attention indexer and main branch from Q8KV8 to Q8KV4; it also adopts Power-aware scheduling to finely allocate compute card power consumption, improving hardware utilization efficiency.
In terms of ecosystem building, MiniMax partnered with chip vendors including Huawei Ascend, T-Head, Kunlunxin, Moore Threads, Biren, Hygon, MetaX, Iluvatar, NVIDIA, and AMD, as well as inference and cloud service platforms including Together AI, Fireworks, Nebius, Baseten, Modal, and DigitalOcean, to enable day-one integration of the model.
In addition, vLLM, SGLang, BAAI FlagOS, as well as Hugging Face and the ModelScope community will jointly advance model adaptation, expanding its application across different chips, cloud platforms, and development environments.
Open-source links:
Hugging Face: https://huggingface.co/MiniMaxAI/MiniMax-M3.1-Flash
ModelScope community: https://modelscope.cn/models/MiniMax/MiniMax-M3.1-Flash