MiniMax publishes the MiniMax M3.1 Flash weights

MiniMax says the weights of MiniMax M3.1 Flash are now public on Hugging Face and ModelScope. When it launched the model on October 19, MiniMax said the weights would follow later.

According to MiniMax:

• Size: 428B parameters in total, 23B activated per token.
• Efficiency: MiniMax Sparse Attention and DSpark speculative decoding improve inference efficiency.
• Benchmarks (MiniMax's own figures, not independently verified): 76.2 on BankerToolBench, 73.7% on DeepSWE v1.1, 61.7 on OSWorld V2, 29.6 on Agents' Last Exam and 82.2 on CyberGym. MiniMax calls this a generational leap in agentic tasks, coding and cybersecurity.
• Other access: the model remains available through the MiniMax API, M Plan and the MCode desktop app.