M3.1 Flash 开源权重规格：428B 总参/23B 激活、1M 上下文、量化后约 250GB
作者拆解 M3.1 Flash 权重：428B 总参、23B 激活、1M 上下文、支持文本图像视频输入；每层稀疏注意力、训练即量化使权重约 250GB，附投机解码草稿头，vLLM/SGLang 首日支持；官方表中 Automation-Bench 62.8 胜过 DeepSeek V4.1 Flash 的 60.5；许
MiniMax M3.1 Flash weights are out. 428B total, 23B active, 1M context, and it takes text, image and video as input. And it's cheap for inference. Sparse attention in every layer. The experts and attention are trained with quantization from the start, so the shipped weights are about 250GB instead of the 800+ a 428B model would take in BF16. Plus, there's a speculative decoding draft head in the repo. vLLM and SGLang support on day 0. In their own table it beats DeepSeek V4.1 Flash 1017 on Automation-Bench, 62.8 to 60.5, and GPT-5.6 Sol and Opus 5 on the same bench by more than 10 points. The license is not MIT. Free for commercial use with a Built with MiniMax credit, and you need their permission above $20M in yearly revenue. For me and for almost every startup that changes nothing. 250GB does not fit my 96GB of VRAM, so I need a smaller quant to run it here. Give the community a few days, it always comes. DeepSeek’s engine on Tuesday, MiniMax weights on Friday. Chinese labs are shipping open weights every few days now. @MiniMax_AI
这条是引用着另一条写的：「MiniMax M3.1 Flash 开源权重发布，Together Compute 首日上线」