In the wee hours of today, Xiaomi made an official announcement regarding the open-sourcing of its MiMo-V2.5 series models. Adopting the MIT license, this series facilitates not only commercial inference deployment but also the retraining process.

The series comprises two models, each capable of supporting a context window of up to 1 million. Specifically, MiMo-V2.5-Pro is tailored for intricate task scenarios and has soared to the pinnacle of global open-source model rankings, as evidenced by its top positions on the GDPVal-AA and ClawEval leaderboards. On the other hand, MiMo-V2.5, a native full-modal model, excels in multimodal understanding and boasts robust Agent capabilities.

Following optimization through public testing, the models within this series have demonstrated remarkable enhancements in both intelligence and stability. With the full release of the model weights, developers are now empowered to freely engage in commercial utilization, fine-tuning, and secondary development. The series initiated public testing on April 23rd, with MiMo-V2.5-Pro proving itself on par with the world's leading Agent models, marking a comprehensive upgrade.