MiniMax Introduces M3 Model for Complex Coding and Automation

MiniMax has introduced its latest AI model, M3, designed for coding agents and automated workflows. According to the South China Morning Post, the Shanghai-based company said the model can process data five times faster than its predecessor while cutting inference costs to one twentieth of previous levels.

Detailed on the company's website, M3 is built on the MiniMax Sparse Attention architecture and supports a context window of up to one million tokens. The model is natively multimodal, trained from the start on text and visual data for deeper alignment across formats.

MiniMax reports that M3 achieved strong results in coding and agent benchmarks, including autonomous task decomposition and multi-step reasoning. In internal tests, it optimized CUDA kernels on Nvidia Hopper GPUs with a 9.4 times speed improvement and replicated an ICLR 2025 paper autonomously over a 12-hour run.

The company plans to release M3 with open weights on platforms such as Hugging Face and GitHub, enabling local deployment and fine-tuning. API users can already access M3 through the MiniMax platform with the same pricing as previous models.