MiniMax plans 2.7-trillion parameter open-source model M3 Pro

MiniMax, the Chinese AI startup behind the popular MiniMax-01 and MiniMax-1.5 models, is preparing to launch its most ambitious model yet: the M3 Pro, a 2.7-trillion-parameter large language model. According to a report from The Information, the internal codename for this model is M3 Pro, and it represents a massive leap from the company's current flagship, the M3, which has 428 billion parameters. The new model is expected to be released and open-sourced as early as the third quarter of 2025, with a focus on dramatically improving complex reasoning and multi-step instruction-following tasks. This move aligns with the broader trend in AI toward scaling model size to enhance emergent capabilities, though it also raises questions about computational cost and accessibility.

By open-sourcing the M3 Pro, MiniMax is following a strategy similar to other Chinese AI players like Alibaba (Qwen) and Baidu (Ernie), aiming to build an ecosystem of developers and researchers around its technology. The sheer scale — 2.7 trillion parameters — would make it one of the largest openly available models in the world, potentially rivaling or surpassing Meta's Llama 4 and Google's Gemini. For professionals, this means access to a model that can handle more nuanced reasoning tasks, longer context dependencies, and complex workflows without the need for costly proprietary APIs. However, deployment will likely require substantial hardware resources, limiting practical use to well-funded teams and cloud providers.