Xiaomi Is Streaming Its RL Run Live and Nobody Else Would Dare
Luo Fuli, who runs Xiaomi's MiMo team and came out of DeepSeek, broke roughly six months of silence on September 16 by doing the single most un-lab-like thing available: turning the training run into a livestream. The dashboard at https://mimo.xiaomi.com/rl streams the reinforcement-learning post-training of MiMo-V2.6 as it happens. Real-time cost, token throughput, benchmark performance, task composition, context length, wall time per step. It hit 533 points on Hacker News, which tells you how starved people are for this.
The numbers on the page are the interesting part. Each step uses 1,568 prompts, 16 rollouts each, roughly 2 billion tokens, all running asynchronously. The task mix is not one domain: code, general reasoning, vision, cybersecurity, and chat are all blended into the same training session, and the run drives multiple harnesses at once. That last detail is the one to sit with. Xiaomi is not training a model against a task, it is training a model against a fleet of environments simultaneously, and you can watch the reward curves argue with each other in real time.
The model is not released. The training run is what is on display. That inversion is the whole point and it is a genuinely aggressive competitive move. Every US frontier lab treats post-training as the crown jewels, the part you describe in a system card six months later in past tense with the numbers sanded off. Xiaomi is publishing the same information while it is still uncertain, which means they are also publishing their failures in advance.
There is a cynical read available and it is probably partly right: a live dashboard is extraordinary recruiting and mindshare for a team that went quiet for half a year and needs to come back loud. Fine. It is still more operational detail about a large RL run than any frontier lab has voluntarily shown, and the fact that a phone company is the one doing it says something uncomfortable about where the openness actually lives in 2026.
The other reason to watch it right now: multiple harnesses in the same run is exactly the variable measured by [the study that found the same model can cost 5x depending on its harness](https://clauday.com/article/402d6e9b-7815-4edb-8115-e41cb6577181). One team is measuring how much the harness costs you at inference, and the other is broadcasting how much it shapes you during training.
← Back to all articles
The numbers on the page are the interesting part. Each step uses 1,568 prompts, 16 rollouts each, roughly 2 billion tokens, all running asynchronously. The task mix is not one domain: code, general reasoning, vision, cybersecurity, and chat are all blended into the same training session, and the run drives multiple harnesses at once. That last detail is the one to sit with. Xiaomi is not training a model against a task, it is training a model against a fleet of environments simultaneously, and you can watch the reward curves argue with each other in real time.
The model is not released. The training run is what is on display. That inversion is the whole point and it is a genuinely aggressive competitive move. Every US frontier lab treats post-training as the crown jewels, the part you describe in a system card six months later in past tense with the numbers sanded off. Xiaomi is publishing the same information while it is still uncertain, which means they are also publishing their failures in advance.
There is a cynical read available and it is probably partly right: a live dashboard is extraordinary recruiting and mindshare for a team that went quiet for half a year and needs to come back loud. Fine. It is still more operational detail about a large RL run than any frontier lab has voluntarily shown, and the fact that a phone company is the one doing it says something uncomfortable about where the openness actually lives in 2026.
The other reason to watch it right now: multiple harnesses in the same run is exactly the variable measured by [the study that found the same model can cost 5x depending on its harness](https://clauday.com/article/402d6e9b-7815-4edb-8115-e41cb6577181). One team is measuring how much the harness costs you at inference, and the other is broadcasting how much it shapes you during training.
Comments