The Pulse
Xiaomi’s MiMo-V2.6 RL Page Offers a Limited Public View
Xiaomi’s MiMo-V2.6 page directs visitors to a live reinforcement-learning project, while a statement from MiMo team lead Fuli Luo describes the run’s scale and planned disclosures.
AI.info Team ·
Xiaomi has published a public page for the reinforcement-learning phase of MiMo-V2.6, giving visitors access to sections labeled “overview,” “metrics” and “about.” The page is presented as a live project site rather than a conventional model card, research paper or product release. Its visible structure points to an ongoing training effort, while a public statement from Xiaomi MiMo team lead Fuli Luo provides additional context about what the project is intended to show.
“MiMo-V2.6 is in the middle of its RL run right now.”
Luo described the work as an active reinforcement-learning run and linked to the Xiaomi page as a stream of the training process. That distinction matters: the page is not a finished report of model capabilities. It is a public window onto work that is still under way, with the eventual model, evaluation results and technical documentation not yet presented as a completed package.
A public MiMo-V2.6 page
The clearest fact established by the official site is that Xiaomi has created a public web presence for MiMo-V2.6 RL. The navigation identifies three areas: an overview, metrics and additional information. The accessible page text does not spell out every field or explain the project in the form of a static technical report, but it does identify the site as a MiMo-V2.6 reinforcement-learning project.
Luo’s accompanying statement says Xiaomi scaled three parts of the work: compute, environments and harnesses, and grader compute. She described roughly 2 billion tokens per step, 1,568 prompts and 16 rollouts, with the process operating asynchronously. She also said the environments combine multiple agentic tasks and harnesses in one run, while grading uses test-case and rubric-based rewards.
Those details come from Luo’s statement rather than from the sparse text exposed in the page’s basic HTML. The distinction should be kept clear when describing the project. The official dashboard is the destination for the live run; Luo’s statement is the source for the explanation of its design and scale.
What the page does and does not establish
The site does not, in the material available for direct verification, provide a conventional account of model parameters, architecture, training hardware or release terms. It also does not present a final benchmark table for MiMo-V2.6. The page’s labels indicate where project information is organized, but they do not by themselves establish the outcome of the run.
Likewise, the page should not be treated as proof of a completed product launch. MiMo-V2.6 is described as still being in reinforcement-learning training. That means figures visible during the run may change, and any evaluation results may not yet represent a final model or final configuration.
Planned disclosures
Luo said Xiaomi would open-source details “piece by piece over the coming weeks.” That promise provides a basis for expecting further technical material, but it is not the same as a release that has already occurred. Until those materials are published, readers should separate the information currently available from later claims about costs, failures, benchmark scores or the final training recipe.
For now, Xiaomi’s page can responsibly be described as a public MiMo-V2.6 reinforcement-learning dashboard, supported by a statement that the run is active and that Xiaomi is scaling compute, agentic environments and grading capacity. It offers more than a product label, but less than a completed model report. A fuller account of the project will require the dashboard’s detailed metrics and the technical documents Xiaomi says it plans to publish.