Xiaomi Opens MiMo-V2.6 RL Training to the Public

A training run becomes a public event

Xiaomi has opened a live dashboard for the reinforcement-learning training of two MiMo-V2.6 models, Flash and Pro. Instead of releasing only weights, code or datasets, the company is exposing parts of the training process itself. Visitors can inspect the number of steps, reward curves, accumulated spending and selected infrastructure events, including GPU failures.

The run is still at an early stage. According to the published figures, the two models had spent more than $1.08 million over roughly 36 hours after Pro training began, or about $30,000 per hour on average. That average should not be interpreted as a fixed future burn rate, but it illustrates the infrastructure cost of running agentic reinforcement learning at scale.

Three ways to scale agentic RL

Xiaomi frames the MiMo system around three scaling dimensions:

• Training compute. Each step processes roughly two billion tokens, with 1,568 prompts and 16 rollouts per prompt. Multiple attempts at the same task give the trainer a set of trajectories to compare rather than a single answer to score.
• Environments and harnesses. MiMo-V2.6 uses multi-task agentic RL, mixing coding, general, visual and chat-oriented tasks in one run. Each task can use a different harness. A coding harness, for example, may let an agent open a terminal, edit files, run tests, inspect errors and continue iterating.
• Grader compute. Complex agent tasks cannot always be judged by a simple success-or-failure label. Xiaomi highlights Agentic In-group Credit Assignment, which seeks to use several rollouts from the same prompt to provide a more detailed signal about which actions helped or hurt. The specific algorithm has not been disclosed.

The pipeline is also fully asynchronous. Rollout generation, environment execution, reward calculation and model updates can proceed at different speeds instead of waiting for a rigid batch sequence. This design is intended to keep a large training system busy while long-running agent tasks are still executing.

Early signals, not a final verdict

The public curves show the Pro model's dynsam/avg@n metric rising from approximately 0.565 to 0.614. Flash increased from about 0.514 to 0.603. In the latest published offline DeepSWE v1.1 evaluation, Pro scored 62.24 and Flash 60.77, compared with 74.2% for DeepSeek-Flash v1.1. Because the run had only been underway for a little over a day, these numbers should not be treated as final rankings. Pro also takes longer to complete a training step, so its latest score was not yet available at the time of reporting.

Why the experiment matters

The main significance is not simply the amount of money being spent. Xiaomi is making visible several normally hidden layers of the RL engineering stack: how trajectories are generated, how tasks enter different environments, how rewards are computed and where the hardware budget goes. The approach suggests that scaling agentic RL requires more than increasing model size. It requires more experiences, more diverse environments and better evaluation signals.

There are also clear limits to what the dashboard can prove. A rising reward curve does not automatically establish stronger general-purpose ability, and the credit-assignment method remains largely unspecified. Independent evaluations will be needed to determine whether the substantial interaction cost can produce reliable transfer beyond the training environments.

Source: QbitAI