Xiaomi Livestreams Training of New Large Language Models

Xiaomi is publicly livestreaming the reinforcement learning training process for its unreleased large language models, MiMo-V2.6-Pro and MiMo-V2.6-Flash. The company has launched a public dashboard that displays metrics directly from the training logs, including reward curves, rollout counts, step timing, and running compute costs.

The dashboard also tracks the models' mid-training coding performance and other evaluation results as the training progresses. Xiaomi has not yet released either model for public use or announced final specifications, pricing, or an API.

Luo Fuli, head of the MiMo team, stated that the team spent nearly six months studying how reinforcement learning could scale after the release of MiMo-V2.5. The dashboard indicates that the MiMo-V2.6 series is "coming soon."

This move towards greater transparency in the model training process is uncommon in the tech industry and provides observers with insights into the development trajectory of large language models.