MiniMax M3 handles 1M tokens, beats GPT-5.5/Gemini 3.1 Pro on coding
Chinese startup's model crushes rivals on coding benchmarks with 5x context boost
MiniMax, a rising Chinese AI startup, has unveiled its new flagship model M3, purpose-built for coding agents and automated workflows. The model boasts a context window of up to 1 million tokens—a fivefold increase over its predecessor—allowing it to handle exceptionally long and complex codebases in a single pass. In benchmark tests on SWE-Bench Pro, a rigorous standard for software engineering tasks, M3 surpassed both OpenAI's GPT-5.5 and Google's Gemini 3.1 Pro, signaling a major leap in coding AI capabilities.
Beyond raw performance, MiniMax announced plans to release M3 under an open-source license with open weights, a move that could accelerate adoption and community-driven improvements. This positions M3 as a potent alternative for developers seeking high-performance, transparent AI models for code generation, debugging, and automated task execution. The combination of massive context and open-source access may disrupt the current AI landscape dominated by proprietary US models.
- Processes up to 1 million tokens of data, 5x more than its predecessor
- Outperforms OpenAI GPT-5.5 and Google Gemini 3.1 Pro on SWE-Bench Pro coding benchmark
- Planned open-source release with open weights for developer use
Why It Matters
MiniMax's M3 challenges US dominance in AI coding, offering an open-source alternative with superior benchmark performance.