MiniMax publishes the M3.1 Flash weights

TechNode reported on October 23, 2026 that MiniMax released the weights for M3.1 Flash that day, for developers to download and deploy. The article says it is a mixture-of-experts model with about 428 billion parameters in total and about 23 billion active per token. Inputs can be text, images, and video, and the context window goes up to 1 million tokens. The service itself launched on October 19. The weights came after that. The image is the title card on the Hugging Face model page, showing the repository name MiniMax-M3.1-Flash.

The model card has no calendar date, so the publication day used here is TechNode's October 23. The card describes a multimodal mixture-of-experts model for coding, agents, computer use, and cybersecurity, with text, image, and video inputs, text output, and a context window of up to 1 million tokens. It says that, relative to MiniMax-M3, this version uses sparse attention throughout, quantization-aware training, and DSpark speculative decoding to improve inference efficiency. Those are design notes on the card, not an independent measurement.

The card's highlight scores are 76.2 on BankerToolBench, 73.7 percent on DeepSWE v1.1, 61.7 on OSWorld V2, 29.6 on Agents' Last Exam, and 82.2 on CyberGym. They are reported here as numbers the card publishes. CyberGym stays a score. The card also says the results are for the official weights and that third-party quantizations may differ. A different parameter count appears in the page sidebar and does not match 428 billion, so it is not used. A date on the sparse-attention paper page is not treated as the date of this release.

TechNode also says the release includes deployment notes for inference frameworks such as SGLang and vLLM, and support for multiple chip platforms through FlagOS. The card recommends those frameworks and gives sampling settings of temperature 1.0 and top_p 0.95. The download command is not repeated here. The card does not say whether 76.2 or 61.7 was reproduced by anyone other than the card's authors. Neither page states a price.