🤖 Xiaomi Streams RL Training of MiMo-V2.6 Models Live
Xiaomi MiMo team lead Luo Fuli (Luo Fuli, previously DeepSeek and Alibaba DAMO) streamed the RL training process of MiMo-V2.6-pro and MiMo-V2.6-flash models live for the first time on September 17, 2026, on the mimo.xiaomi.com/rl page. Training is fully asynchronous: at each step, 1568 prompts with 16 rollouts are taken — about 25,000 attempts and approximately 2 billion tokens. At the time of checking, the dashboard showed step 17 for flash with a passrate of 0.585 on mixed code/general/chat/visual domains. Costs have already exceeded $1.1 million: about $806,000 for pro over 12 steps, $354,000 for flash over 17 steps — approximately $30,900 per hour. In the DeepSWE v1.1 benchmark, pro shows 63.72, flash — 60.77. Documentation and developments are promised to be opened in the coming weeks; the incident log shows restarts of both runs due to network and VRAM failures, and the cybersecurity dataset was removed from the upcoming pro run due to "bad patterns" in rollouts.
🌍 For the first time, the company is streaming telemetry of a frontier RL run in real time — usually training is closed until the model release. Publicly visible are verifiable RL-scaling mechanisms: token volume per step, task distribution across domains, rewards, and infrastructure incidents. The experiment "to what limits RL scales" has become observable, not just claimed in a press release: numbers that are usually disclosed in a technical report months later are read live from the dashboard here.
👤 The mimo.xiaomi.com/rl dashboard can be opened right now to watch steps, passrate, and timings live, and the promised opening of documentation will provide specific recipes for asynchronous multi-agent RL, comparable to DeepSeek and OpenAI approaches. The open incident log is also instructive: it shows what real failures of large RL runs look like. However, a passrate of 0.585 by itself is not an indicator of progress — without a definition of the reward and domain weights, this number is incomparable neither between steps nor with other runs.
Source 1: https://x.com/_LuoFuli/status/2100296686719610932 Source 2: https://mimo.xiaomi.com/rl/#overview