MiniMax-M3 is the flagship frontier model from MiniMax, designed to bring together frontier-level coding, native multimodality, and extremely long context handling in a single open-weight system. The model introduces MiniMax Sparse Attention (MSA), an architectural approach specifically engineered to handle the computational demands of very long contexts without degrading quality. This design makes it well-suited for developers who need to reason across large, mixed-format inputs—simultaneously processing video, images, and text—while maintaining coherent agentic behavior. The sparse attention mechanism is what enables the million-token context window to remain practical, addressing a common bottleneck in long-document and multi-file coding tasks.

MiniMax-M3 shows its practical strength in agentic reasoning and automation workflows. On benchmarks, it scores 59.0% on SWE-Bench Pro and 83.5 on BrowseComp, with MiniMax asserting it surpasses GPT-5.5 and Gemini 3.1 Pro on coding while edging past Claude Opus 4.7 on autonomous browsing tasks. These results reflect a model built for real-world developer workflows: navigating repositories, executing terminal commands, and reasoning across large codebases. The open-weight approach lowers the barrier for teams wanting to fine-tune or deploy specialized agents, positioning M3 as a foundation for coding assistants, automated tooling, and research pipelines that require both depth and breadth across modalities.