Xiaomi Shares Live Training Data for MiMo-V2.6 Pro and Flash
Xiaomi is publishing live reinforcement-learning training metrics for mimo-v2.6-pro and mimo-v2.6-flash, pulled directly from trainer logs. The tracker at mimo.xiaomi.com/rl went live on September 15, the same date the MiMo-V2.6-Pro run started. It reports metrics from both RL runs as training proceeds. MiMo RL tracker
The page covers the two current branches, mimo-v2.6-pro and mimo-v2.6-flash. Xiaomi says the feed comes straight from trainer logs, the automatic records generated during training, with no summarized or delayed release. That makes the training process itself visible while it is still underway.
The MiMo-V2.6-Pro run began on September 15, 2026. It was consuming computing resources at a rate of approximately $432,000 per day, according to reporting published on September 17. Yahoo Tech report That figure is the daily burn rate during the run, not a total training cost.
Xiaomi positions MiMo-V2-Pro as its flagship foundation model "built for real-world agentic workloads." The dedicated product page describes the model as "designed to serve as the brain of agent systems." MiMo-V2-Pro page The wording is consistent across the page. It puts agency, software that can use tools and carry out multi-step jobs, ahead of chat or single-turn answers as the main design goal.
MiMo-V2-Flash sits alongside it with a different focus. Xiaomi calls it a "powerful, efficient, and ultra-fast foundation language model that excels in reasoning, coding, and agentic scenarios." MiMo-V2-Flash blog In technical terms, it is a Mixture-of-Experts model, meaning it holds many specialist submodels but only activates a few for each step, with 309B total parameters and 15B active parameters.
The immediate predecessor in the public lineage is MiMo-V2.5. It is a 310B-parameter sparse Mixture-of-Experts model with 15B active parameters trained on 48T tokens. MiMo-V2.5 page Its context length, the amount of text it can consider at once, was extended from 32K to 256K to 1M during development.
A larger sibling followed. Xiaomi released mimo-v2.5-pro on April 23, 2026, with 1T total parameters and 42B activated parameters. The documentation lists that scale without additional training details in the cited update log. Together, the public numbers show a 15B active-parameter level for V2.5 and Flash, and a step up to 42B active for v2.5-pro.
The broader context here is transparency as an engineering signal. Live trainer-log data does not change the underlying training task. It does change what outside practitioners can learn from it. Training curves, stalls, restarts and steady-state behavior become visible while decisions are still being made, which is more useful to other RL practitioners than a polished chart shared afterward. In my view, that is the point worth watching.
Looking at what this means for agent work, continuity matters more than any single run. Pro is framed around agentic workloads, Flash pairs reasoning and coding with speed, V2.5 extended context by more than an order of magnitude, and v2.6 now shows RL work in public. For teams building tool use, multi-step planning and code-driven agents, longer context plus RL-tuned behavior is a practical combination. It lowers the cost of keeping state in context and raises the chance the model follows through across steps. The feed will be most useful if Xiaomi keeps it running through the noisy middle of RL, not only the early gains.
