MiniMax publishes open weights for M3.1 Flash, a 428B parameter model with 23B active per token, citing benchmark scores that are its own figures
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 82.2 on CyberGym. Efficient by design: MiniMax Sparse Attention and DSpark speculative decoding improve inference efficiency. Hugging Face: t.co/JDYA1zZh54 ModelScope: t.co/DQcmOjTwhO API: t.co/fHRdSV7BwZ M Plan: t.co/LYnYZcsesu MCode for desktop:
Context
Downloadable weights for a large multimodal coding and agent model let outside teams run and tune it themselves, but every performance number so far comes from MiniMax and has not been independently checked.
MiniMax posted at 9:06 PM Eastern on October 22 that M3.1 Flash weights are public, with links to Hugging Face and ModelScope. The Hugging Face model card describes a native multimodal mixture of experts model with a context window of up to 1M tokens. The post is a company announcement, and the benchmark results in it are MiniMax's own claims. The post quotes MiniMax's earlier announcement of the model.
Deep dive
MiniMax said on X that the weights for MiniMax M3.1 Flash are now public, with 428B parameters in total and 23B activated per token. The post lists 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. These are MiniMax's own figures. It also links to its API, its M Plan and MCode for desktop.
The Hugging Face model card describes the model as built for coding, agentic workflows, computer use and cybersecurity. It accepts text, image and video input, produces text, and supports up to 1M tokens of context. According to the card, M3.1 Flash builds on MiniMax M3 with sparse attention throughout the model, quantization aware training and DSpark speculative decoding, and the release includes the DSpark draft head.
The card's comparison table sets MiniMax's own results next to DeepSeek V4.1 Flash, Kimi K3, GLM 5.3, GPT 5.6 Sol and Opus 5. On several rows the table shows at least one rival scoring higher than M3.1 Flash, for example DeepSWE v1.1 (77.0 for DeepSeek V4.1 Flash 1017 against 73.7) and OSWorld 2.0 (65.7 for GPT 5.6 Sol and 70.6 for Opus 5 against 61.7). The card notes that results are for the official weights and that third party quantizations may differ. Its evaluation notes list the agent harness versions MiniMax used.
The card recommends SGLang, vLLM and TokenSpeed for serving. The Hugging Face page's automatic size panel shows 220B parameters, which differs from the 428B in the card and post; the page does not explain why.
What to expect next
The card lists integrations for Transformers, KTransformers, unsloth and ATOM as still in progress. Independent evaluations of MiniMax's benchmark claims had not been published in the sources reviewed. The card says third party quantizations may behave differently from the official weights, so early community reports on smaller builds may not match the figures above. The licence terms for commercial use were not reviewed here and should be read on the model page before deployment. MiniMax lists model@minimax.io as its contact for the release.