Xiaomi Live-Streams MiMo-V2.6 RL Training, Burning About $30,000 an Hour

The training is still at an early stage, just over one day in, but Xiaomi's page already shows gains from RL. Pro's dynsam/avg@n rose from about 0.565 at step 1 to 0.614, while Flash rose from about 0.514 to 0.603. In the latest public DeepSWE v1.1 offline evaluation, Pro and Flash reached 62.24 and 60.77, respectively. For comparison, DeepSeek-Flash v1.1 scored 74.2%. Flash has closed much of the gap from a lower starting point, while Pro retains a small lead. Pro takes longer to complete one step, so its latest score has not yet appeared.

After open-sourcing MiMo-v2.5 in April, Xiaomi stayed quiet for nearly half a year. Luo said the team studied only one question during that period: how far reinforcement learning can scale. Pretraining scaling has a familiar formula, more data, more parameters and more compute leading to stronger models. Frontier labs are now asking whether RL can scale in a similar way by increasing the number of trajectories generated per round, the number of real task environments the model faces, the compute spent judging whether those trajectories are good, and whether model capability can keep improving.

Xiaomi's answer is to scale along three dimensions: training compute, environments and harnesses, and grader compute. For training compute, each MiMo-V2.6 training step processes about 2 billion tokens, configured as 1,568 prompts with 16 rollouts per prompt. That means the same problem is not answered once but explored through many different solving and action trajectories, producing more than 25,000 rollouts in a single round. For agent tasks, one rollout is far more than a question and answer. The model may go through dozens of rounds of thinking, tool calls, environment feedback and further actions, so a single step can reach billions of tokens.

The system uses a fully asynchronous execution model. Traditional RL training can be imagined as an orderly pipeline: generate all samples, score them, train the model, then start the next round. Large-scale agent RL struggles with that kind of waiting. In MiMo's system, different tasks can be at different stages at the same time: some agents are still executing in an environment, some have finished and are waiting for scoring, some are having rewards computed, and new tasks are entering the system. Generation, execution, scoring and training no longer queue strictly; they form a continuously running asynchronous pipeline.

The second scaling direction is environments and harnesses. MiMo-V2.6 runs multi-task agentic RL, mixing code, general tasks, vision and chat tasks into the same RL run, and those tasks can run in different harnesses. A coding agent's harness may let the model open a terminal, modify code, run tests, read errors and continue editing. A vision agent may face a different set of tools, environment feedback and success conditions. Xiaomi therefore wants to scale not only the number of problems but also the number of environments, tools and problem types the model can enter. The live page displays batch composition, showing what kinds of tasks and environments the model is learning from in each training step.

The third direction is grader compute, or adding compute to scoring itself. Reinforcement learning needs rewards, but complex agent tasks do not produce rewards as easily as a math multiple-choice question. Code tasks can run test cases, and the number of tests passed out of 100 can provide relatively objective feedback. For more open agent tasks, judging only final success or failure is often far from enough, so the grader that decides how well the model performed consumes increasing amounts of compute.

A more critical issue is credit assignment. Suppose an agent executes 40 steps to complete a task: searching for material, opening webpages, reading information, writing code, running tests, finding an error, modifying code and finally succeeding. If the system only tells the model success, reward = 1, the learning signal is very coarse. The model does not know which of the 40 actions actually helped, which steps were unnecessary, which edit turned the task around, or which actions did not cause failure but were poor choices. MiMo mentions agentic in-group credit assignment. No specific algorithm has been published, but with the same prompt generating 16 rollouts, the basic goal can be understood as using multiple agent trajectories and their outcomes from the same group to obtain a finer reinforcement learning signal than simple final success or failure. The system must decide which behaviors should be encouraged more, which less, and how to turn that judgment into a training signal.

Xiaomi's three scaling directions form a single system. Scaling compute lets the model produce more experience. Scaling environments and harnesses lets it gain more diverse experience. More grader compute extracts more accurate learning signals from that mass of experience. After years of pretraining scaling, the question is whether the entire RL loop, in which a model interacts with environments, produces experience, evaluates it and learns from it, can also become a machine that keeps growing in scale. A prominent ViT researcher who watched the stream offered a plainer translation: three things, all backed by compute.