- AI
- LLM
- Open source
MiniMax opens the weights of M3.1 Flash
MiniMax has released the weights of M3.1 Flash, a mixture-of-experts model with 428 billion total parameters and 23 billion activated per token, as a public, downloadable model on Hugging Face and ModelScope.
·1 min read·AB
MiniMax (official) @MiniMax_AI
MiniMax M3.1 Flash weights are now public. 428B parameters total, 23B activated per token.
Agentic performance: A generational leap in agentic tasks, coding, and cybersecurity — 76.2 on BankerToolBench, 73.7% on DeepSWE v1.1, 61.7 on OSWorld V2, 29.6 on Agents' Last Exam, and… https://t.co/0sIsMoqIp7
MiniMax has released the weights of M3.1 Flash, a mixture-of-experts model with 428 billion total parameters and 23 billion activated per token, as a public, downloadable model on Hugging Face and ModelScope. It is also available through MiniMax's own API and desktop agent tool, MCode.
According to MiniMax, the model takes a generational leap in agentic tasks, coding, and cybersecurity, citing its own benchmark scores of 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 says the model uses its own Sparse Attention mechanism throughout, plus a technique called DSpark speculative decoding, to improve inference efficiency. Together AI, which lists the model with a 1M token context window and support for text, image, and video input, notes it builds on an earlier MiniMax model called M3.
MiniMax also thanked infrastructure partners Inferact and SGLang for supporting the model from launch day, with SGLang saying the model can decode at more than 1,000 tokens per second in certain configurations. All benchmark and performance figures here come from MiniMax and its partners.
Why it matters
Open weights for a large, benchmark-competitive agentic model give developers a free alternative to closed coding and tool-use models.
Source: MiniMax