Shanghai-based MiniMax AI has claimed its new flagship model, M3, can outperform OpenAI’s GPT-5.5 and Google’s Gemini 3.1 Pro on the SWE-Bench Pro coding benchmark, though the results have been reported by MiniMax and require independent testing. MiniMax says M3 is a coding-focused model for software engineering tasks, featuring a one million-token context window to process large volumes of text or code. The company is positioning M3 for AI agents and automated workflows, not just one-off code generation. MiniMax has described M3 as using a redesigned architecture called MiniMax Sparse Attention, which the company says focuses on the most relevant information inside large datasets to reduce computing requirements and improve speed when working with long contexts.