MiniMax-M3

MiniMax M3 is a multimodal foundation model that processes text, images, and video inputs while generating text output, designed specifically for sustained, multi-step development tasks rather than quick single-turn queries. Its defining architectural choice is MiniMax Sparse Attention (MSA), which replaces traditional full attention with a selective KV-block mechanism—a lightweight index branch scans incoming tokens and routes computation only to the most relevant key-value blocks. This approach achieves roughly one-twentieth the computational cost of the previous generation when operating at 1M tokens, delivering substantially faster prefill and decode without sacrificing output quality. The 1M-token context window makes it well-suited for reading large codebases, debugging complex projects, generating files across an entire repository, and handling intricate software development workflows that demand sustained attention across many files.

The model was trained as a natively multimodal system on interleaved data, meaning it learned to reason across text, images, and video from the ground up rather than bolting on vision capabilities afterward. Training incorporated an interactive user-simulator framework that tuned the model for multi-turn, production-like collaboration—essentially teaching it to maintain coherent reasoning across extended conversations. Reviewers have noted this generation marks a meaningful step forward from earlier M-series models, with one analyst suggesting the capability gap compared to leading proprietary models may have narrowed. The model is available with open weights, allowing developers to run it locally or through various hosted providers, and its combination of long-context capacity, multimodal understanding, and agentic reasoning positioning it for tooling ecosystems where sustained autonomous problem-solving matters more than isolated responses.