en.Wedoany.com Reported - MiniMax open-sourced the MiniMax M3 model weights last Friday (June 12) and simultaneously released the MSA (MiniMax Sparse Attention) technical paper. The model has been officially open-sourced recently.

M3 is MiniMax's native multimodal flagship model, with a total of 428B parameters and 23B activated parameters. As the first open-source model to undergo multimodal hybrid training from scratch, M3 has achieved the highest ranking among global open-source models on the Artificial Analysis Comprehensive Intelligence Index within two weeks of its release.

In response to user experience issues caused by increased traffic, the company stated that it is continuously optimizing. The output speed of the M3 model has been improved from approximately 30 TPS at launch to about 80 TPS, with plans to further increase it by 30-40% to enhance response efficiency. In coding and agent evaluations, M3 has demonstrated capabilities in autonomous task decomposition, tool invocation, and multi-step reasoning, with the quality of its generated code described as production-ready. The open-source code for this model has been released on the Hugging Face platform at the link: https://huggingface.co/MiniMaxAI/MiniMax-M3.

This article is compiled by Wedoany. All AI citations must indicate the source as "Wedoany". If there is any infringement or other issues, please notify us promptly, and we will modify or delete it accordingly. Email: news@wedoany.com