MiniMax releases weights for 428-billion-parameter M3.1 Flash model
Shanghai-based AI firm MiniMax has released the weights for its M3.1 Flash model, a 428-billion-parameter architecture capable of processing text, images, and video.
MiniMax has made the weights for its M3.1 Flash model available for public download, enabling developers to integrate the large-scale model into local and cloud-based applications. The model features approximately 428 billion total parameters, with a sparse activation of 23 billion parameters per token, designed to balance capacity with operational efficiency.
The model supports a context window of up to 1 million tokens and is optimized for multimodal tasks including image and video processing, as well as complex coding and autonomous agent workflows. Technical optimizations include sparse attention mechanisms, quantization-aware training, and speculative decoding to reduce inference latency.
To facilitate adoption, the release includes integration support for inference frameworks like SGLang and vLLM. Furthermore, the model is compatible with various hardware platforms via the company's FlagOS. This release follows the initial deployment of the model as a service on October 19.