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

Summary

MiniMax released the weights of its M3.1 Flash model as open source. This 428-billion-parameter Mixture-of-Experts (MoE) architecture activates only 23 billion parameters per token; it supports text, image, and video inputs and offers a 1-million-token context window.

Why it matters

Releasing a model of this scale with open weights is a significant step toward democratizing frontier-class AI. In particular, the MoE design keeps per-token computational costs much lower compared to a dense 428B-parameter model. Multimodal inputs, a 1-million-token context, and a clear focus on coding, agent, and computer-use workloads position the model as a direct competitor to closed-source alternatives. The inclusion of deployment tools (SGLang, vLLM, FlagOS multi-chip support) lowers the practical barrier for teams looking to host it on their own infrastructure.

Highlights

• 428 billion total parameters / 23 billion active parameters per token thanks to MoE; sparse attention mechanisms, quantization-aware training, and speculative decoding for efficiency
• Multimodal (text + image + video), up to 1 million token context; explicitly targeted at coding, AI agents, and computer-use tasks
• Weights were released following the hosted service launch on October 19; deployment guides for SGLang, vLLM, and FlagOS across multiple chip platforms were provided

Critical take

A reasonable, two-paragraph press release recap. There is no benchmark data; no head-to-head comparison was made with similar open or closed models. The license terms governing the weights are never mentioned; yet for a developer considering adoption, the license is undoubtedly the most critical detail. The four-day gap between the hosted service launch (October 19) and the release of the weights (October 23) is unexplained. The claim that the model is 'designed for coding, AI agents, and computer-use tasks' is made without a single supporting data point.