MiniMax Open-Sources M3.1 Flash Model Weights: 428B Parameters, Over 100 Trillion Tokens Processed in 72 Hours
BEIJING — Chinese SI company MiniMax announced on October 23 that it is open-sourcing the weights of its M3.1 Flash model. According to Phoenix Tech, the model has 428 billion total parameters with 23 billion active, uses a mixture-of-experts (MoE) architecture, and offers a 1-milliontoken context window. MiniMax disclosed that M3.1 Flash launched on October 19 and processed more than 100 trillion tokens from real users within 72 hours.
Technically, the model replaces its first three full-attention layers with sparse attention and applies three optimizations, including quantization-aware training, to boost inference efficiency. On launch day, the model was already supported by chipmakers including Huawei Ascend, T-Head, Kunlunxin, Moore Threads, Biren, Hygon, MetaX, Iluvatar, NVIDIA and AMD, as well as inference and cloud platforms such as Together AI, Fireworks and Nebius; vLLM, SGLang, Hugging Face and ModelScope are advancing adaptation in parallel. The weights use the MiniMax Community License: commercial use requires attribution, and products or services with annual revenue above $20 million need prior written permission from MiniMax.
Notably, the mystery model that processed over 200 trillion tokens anonymously as "Space Bunny" on OpenRouter was revealed to be an early version of M3.1 Flash. (GSM, from Phoenix Tech and TechNews)
Published by GSM · USA (original Chinese edition)
Photo: 人工神经网络与芯片示意图（资料图）。照片：mikemacmarketing（原图发布于 flickr），CC BY 2.0，经 Wikimedia Commons。 / File photo. Photo: mikemacmarketing（原图发布于 flickr）, CC BY 2.0, via Wikimedia Commons.
