ChainThink reports that, on July 8, according to The Information citing sources, MiniMax is developing a 2.7-trillion-parameter large language model, which could be released as early as the third quarter and is planned to be open-sourced.
The internal code name for this model is M3 Pro, and the final release name has not yet been determined. If the weights are opened as planned, this model is expected to become the largest open-source large model in the world by parameter count.
MiniMax’s current flagship model, M3, has 428 billion parameters; the new model is approximately 6.3 times larger and will surpass several existing trillion-parameter open-weight models.
The report also noted that larger parameter scales generally benefit complex reasoning and multi-step instruction tasks, but the model's final performance still depends on training quality, activated parameters, inference costs, and evaluation results.