Xiaomi released an updated language model MiMo v2.6

Company Xiaomi introduced an updated version of its language model MiMo v2.6. The release is aimed at increasing the transparency of algorithm work and reducing the cost of operation for end users.

During testing in the Terminal Bench 4.0 benchmark, the model scored 141 points. The training and fine-tuning process took only 1 hour, indicating the high efficiency of the architecture when working with computational resources. Developers emphasize that the model demonstrates stable performance indicators when performing standard natural language processing tasks.

Users have access to two main variations of the system: Flash and Pro. Both versions of the model are already placed on the HuggingFace platform, allowing developers to integrate them into third-party projects. Technical documentation and weight specifications are available for review in the project's official repository.

This release continues the company's strategy of developing accessible AI tools. The MiMo v2.6 update focuses on optimizing weights, allowing the model to be launched on less powerful hardware without significant loss of answer accuracy. In the Hacker News community, a discussion of the architectural features of the new version and its comparison with previous iterations of the MiMo line has already begun.