MiniMax Debuts M3: Flagship AI Model for Complex Coding Tasks

Summary

Shanghai-based AI startup MiniMax has announced M3, a new flagship AI model designed to power coding agents and automated workflows. The model represents a significant leap forward in capability, capable of processing up to 1 million tokens at once—five times more than its predecessor, the M2.7—enabling it to handle long, complex programming projects. In benchmark tests, M3 demonstrated superior performance on SWE-Bench Pro, a major coding benchmark, outperforming OpenAI's GPT-5.5 and Google's Gemini 3.1 Pro.

MiniMax achieved these gains through architectural redesign that dramatically reduced computational requirements to as little as one-twentieth of previous levels. This efficiency improvement translates directly to faster inference speeds and significantly lower costs, addressing two key pain points for AI practitioners deploying models at scale. The company demonstrated M3's capabilities by successfully optimizing software running on Nvidia's Hopper chips, showcasing its ability to tackle real-world engineering challenges.

• Designed specifically for coding agents and automated workflows, positioning MiniMax in the rapidly growing AI automation market
• 5x faster processing compared to M2.7 predecessor while delivering superior benchmark performance

Editorial Opinion

MiniMax's M3 launch signals intensifying competition in the specialized coding AI market, where both established giants and nimble Chinese startups are racing to dominate AI-assisted software engineering. The model's superior SWE-Bench Pro performance and architectural efficiency gains demonstrate that specialized, well-optimized models can compete effectively against larger, more generalized counterparts from OpenAI and Google. If MiniMax can effectively commercialize M3 for the global developer market, it could establish the company as a serious contender in the enterprise AI space.