Chinese AI startup MiniMax has unveiled its new flagship model, the M3, which should become the basis for the company's entry into the market of programming agents and automated workflows.
The company said that the updated model architecture reduced computational costs by 20 times, which reduced inference costs and significantly accelerated response times.
Chinese AI startup MiniMax has unveiled its new flagship model, the M3, which should become the basis for the company's entry into the market of programming agents and automated workflows.
The company said that the updated model architecture reduced computational costs by 20 times, which reduced inference costs and significantly accelerated response times.
MiniMax noted that the M3 is capable of simultaneously processing up to 1 million data tokens, which is five times more than its predecessor, the M2.7. This volume allows the model to work with long and complex programming projects. In one of the benchmarks provided by the company, the M3 successfully solved the task of optimizing software running on Nvidia Hopper chips, reports the South China Morning Post.
According to the company, MiniMax-M3 outperformed competitors such as OpenAI’s GPT-5.5 and Google’s Gemini 3.1 Pro in the well-known SWE-Bench Pro coding test. This may indicate the model’s high efficiency in solving software engineering and process automation tasks.
The company did not disclose the size of the model or the computing infrastructure it used to train it.
