MiniMax debuts AI model built for long and complex coding tasks
MiniMax, a Shanghai-based AI startup, unveiled its latest model M3, which processes up to 1 million tokens of data at once—five times more than its predecessor—and outperformed OpenAI’s GPT-5.5 and Google’s Gemini 3.1 Pro on coding benchmarks. The model’s redesigned architecture reduces computational costs to one-twentieth of previous levels while preparing for an IPO on Shanghai’s Star Market.
MiniMax, a Shanghai-based AI company, introduced its new model M3 on Monday, targeting automated coding and workflow tasks. The model’s architecture reduces computational demands to as little as one-twentieth of prior versions, cutting inference costs while improving speed. M3 can handle up to 1 million tokens of data, five times more than its predecessor M2.7, enabling it to tackle complex programming projects. In benchmark tests, M3 outperformed competitors like OpenAI’s GPT-5.5 and Google’s Gemini 3.1 Pro on the SWE-Bench Pro coding benchmark, demonstrating strong capabilities in software engineering. MiniMax demonstrated M3’s ability to optimize software for Nvidia’s Hopper chips in a test case. The company did not disclose the model’s size or training infrastructure but highlighted its efficiency gains. MiniMax plans to list M3 as part of its preparations for an initial public offering on Shanghai’s tech-focused Star Market, complementing its existing Hong Kong listing. The model’s focus on long, complex coding tasks positions MiniMax as a key player in AI-driven automation. Its WeChat announcement emphasized M3’s cost-effectiveness and performance improvements over prior iterations.
This content was automatically generated and/or translated by AI. It may contain inaccuracies. Please refer to the original sources for verification.