MiniMax releases open weights for 428-billion-parameter M3.1 Flash model

MiniMax's release of the "weights" for its M3.1 Flash model on October 23 means developers can now download and deploy the model on their own infrastructure, rather than relying solely on MiniMax's hosted service. The M3.1 Flash is described as a mixture-of-experts (MoE) model, employing a design where only a subset of its 428 billion total parameters, specifically 23 billion, is activated per token. This approach is typically used to enhance inference efficiency, which refers to the speed and cost of running the model to generate outputs.

The model is reportedly designed for coding, AI agents, and general computer-use tasks. MiniMax states it improves inference efficiency through techniques like sparse attention, quantization-aware training, and speculative decoding. These methods are established ways to optimize how large models process information, aiming to reduce computational demands without necessarily compromising output quality.

To facilitate deployment, MiniMax has included guidance for popular inference frameworks such as SGLang and vLLM. Additionally, the release provides support for multiple chip platforms through FlagOS. This practical support suggests an effort to reduce the technical barriers for developers looking to integrate and run a large, multimodal model with a focus on operational efficiency within their own environments.