The Pulse
MiniMax Publishes M3.1 Flash Weights Under Community License
MiniMax’s M3.1 Flash weights are available on Hugging Face under a community license with conditions for commercial use. The repository reports about 220.5 billion parameters in its metadata, while its model-card text separately describes a
AI.info Team · Written with AI · fact-checked ·
MiniMax has made M3.1 Flash’s weights available for download under a community license that sets conditions for commercial use. The Hugging Face repository’s metadata reports 220,490,854,016 parameters, or about 220.5 billion. Its separate model-card text says approximately 428 billion total parameters, with 23 billion activated per token; the page does not explain the difference. The repository’s file metadata totals 250,127,284,237 bytes.
Downloadable weights, company-set terms
The license allows noncommercial use without charge. For commercial products or services using the software, it requires prominent attribution and a one-time email notice to MiniMax. Products and services generating more than $20 million in yearly revenue must instead obtain prior written authorization. The license defines commercial use to include paid services, commercial use of APIs, and commercial deployment of modified or fine-tuned versions.
The required attribution is “Built with MiniMax M3.1-Flash,” which the license says must appear on a related website, user interface, blog post, about page or product documentation. The repository provides instructions for running the model with vLLM and SGLang. Its listed size—about 250.1 billion bytes—means users should account for the storage and serving resources needed to download and run the weights.
Multimodal inputs and a long context window
MiniMax describes M3.1 Flash as a multimodal mixture-of-experts model for coding, agentic workflows, computer use and cybersecurity. It accepts text, image and video inputs, produces text, and supports a context window of up to 1 million tokens. The model card cautions that usable context length and throughput depend on GPU memory and serving configuration, so the maximum context figure does not guarantee the same performance across deployments.
The model card says M3.1 Flash builds on MiniMax M3 with sparse attention throughout, quantization-aware training and DSpark speculative decoding. It also says the release includes a DSpark draft head for compatible serving implementations. MiniMax’s October 19 announcement said the model was available through its API, M Plan and MiniMax Code, while its weights would be released later. The Hugging Face repository now hosts the downloadable files.
Published benchmark results
The model card lists results of 76.2 on BankerToolBench, 73.7% on DeepSWE v1.1, 61.7 on OSWorld 2.0 and 82.2 on CyberGym. Its comparison table lists DeepSeek V4.1 Flash-1017 at 77.0 on DeepSWE, Opus 5 at 70.6 on OSWorld 2.0, and GLM-5.3 at 84.5 on CyberGym. On BankerToolBench, the table gives M3.1 Flash and Opus 5 the same score, 76.2.
The repository says results are for the official weights and warns that third-party quantizations may differ. The comparisons are MiniMax’s published figures; teams evaluating the model for their own work can test it against their own tasks and serving setup.
The license is part of the deployment decision
Making the weights public allows developers to download and run M3.1 Flash outside MiniMax’s hosted products, subject to the license. Commercial users must account for the attribution requirement and notice to MiniMax, while products and services above the $20 million yearly-revenue threshold require written authorization. The repository’s parameter-count discrepancy, file size and hardware-dependent context limit are also relevant details for organizations assessing a deployment.