MiniMax published the weights of its M3.1 Flash model on October 23, allowing developers to download and implement this model. This mixture-of-experts type model has approximately 428 billion total parameters, with 23 billion parameters activated for each token.
M3.1 Flash supports processing text, image, and video inputs, and features a context window reaching one million tokens.
The model was developed to perform tasks related to coding, AI agent operation, and general computer operations. To improve inference efficiency, M3.1 Flash utilizes sparse attention methods, quantization-aware training, and speculative decoding.
The release includes deployment guides for inference frameworks such as SGLang and vLLM, as well as support for multiple chip platforms via FlagOS. These weights were released following the initial launch of the model service on October 19.