Xiaomi has streamed the RL training process for the MiMo-V2.6-pro and MiMo-V2.6-flash models publicly for the first time: since September 17, 2026, run telemetry has been available in real time on the mimo.xiaomi.com/rl page. Each step of the asynchronous training executes about 25,000 rollouts and processes roughly 2 billion tokens, total costs have exceeded $1.1 million, and the open dashboard shows passrate, task distribution across domains, and a failure log. In the coming weeks, the team promises to publish documentation and findings, which could give engineers the first open recipe for asynchronous multi-agent RL.

What happened

The stream was launched by Xiaomi MiMo team lead Luo Fuli on September 17, 2026, on the mimo.xiaomi.com/rl page. Training for both models is fully asynchronous: each step takes 1,568 prompts with 16 rollouts each, yielding about 25,000 attempts and roughly 2 billion tokens per step. At the time of checking, the dashboard showed step 17 for flash with a passrate of 0.585 and mixed code/general/chat/visual domains; for pro, 12 steps had passed by that point. Costs exceeded $1.1 million: about $806,000 spent on pro over 12 steps and about $354,000 on flash over 17 steps at a rate of roughly $30,900 per hour. In intermediate DeepSWE v1.1 benchmark measurements, pro shows 63.72 and flash — 60.77. The open incident log records restarts of both runs due to network and VRAM failures, and a cybersecurity dataset was removed from the nearest pro run due to "bad patterns" in rollouts.

Context

Usually, the RL training stage of frontier models is closed until release: external observers only see the final technical report with selected numbers, and the live dynamics of the run remain internal lab information. Xiaomi is taking the opposite path, turning training into an observable experiment that can be verified in real time rather than taken on faith from a press release. The run setup is unusual: a single run mixes code, general, chat, and visual domains and simultaneously uses multiple test harnesses — effectively a claim to a multi-domain asynchronous RL recipe, the value of which will only become verifiable after the methodology is published. The story is further weighted by Luo Fuli's background: he leads the MiMo team after working at DeepSeek and Alibaba DAMO, meaning the person behind the stream is familiar with the closed practices of leading labs. The quality benchmark at the intermediate stage is the DeepSWE v1.1 benchmark, and the promised open recipes can be directly compared with the approaches of DeepSeek and OpenAI.

Why this matters for the industry

For the industry, this is the first known case of a company streaming frontier RL run telemetry in real time: the RL-scaling mechanism becomes observable rather than claimed in a post-release technical report. Engineers and researchers have access to rare market data on the economics of a large run — about $30,900 per hour and over $1.1 million by mid-run — which can be used to plan their own RL budgets and capacity calculations. The open incident log with restarts due to network and VRAM failures shows the real cost of reliability for large runs, which is usually hidden. For startups, the signal is twofold: the promised documentation on asynchronous multi-agent RL could lower the barrier to entry for RL scaling and provide a blueprint for comparison with the approaches of DeepSeek and OpenAI, but at the same time, transparency works as a GTM move, raising the bar of expectations before the model release. If the practice of public telemetry takes hold, the industry norm may shift from "technical report after release" to real-time observable training runs, and the accumulated passrate curves, failure logs, and costs per step will become a body of data for studying the dynamics and economics of RL scaling.

Why this matters for users

The mimo.xiaomi.com/rl dashboard is open right now: you can observe training steps, passrate, task distribution across domains, timings, and the failure log live — without waiting for a release and without paid access. For engineers, this provides a free case for incident management exercises and calibrating expectations on the speed and cost of RL training, and for readers — a rare opportunity to see how large runs actually go, including failures and restarts that companies usually don't show. Practical limitation: there is currently nothing to deploy MiMo-V2.6 in products — there is no API access, pricing, or latency metrics in the sources, and the models are still training, so the value right now is in observation and telemetry accumulation, not integration. For analysts and technical content authors, live numbers provide material for comparative breakdowns and forecasts before the models are released.

What is still unknown / limitations

The passrate of 0.585 at step 17 for flash cannot be read as a progress indicator: without a definition of the reward, domain weights, and a breakdown by code/general/chat/visual, this number is not comparable across steps, across domains, or with other runs. The numbers 63.72 and 60.77 in DeepSWE v1.1 are intermediate measurements of models still in training, not a final evaluation. The removal of the cybersecurity dataset from the pro run is a fact from the source, but interpreting it as a signal of rollout quality instability when mixing domains requires confirmation after the methodology is disclosed. The promise to open up documentation and findings has not yet been fulfilled; if only part of the materials is opened, conclusions about the reproducibility of the recipe will have to be limited. Finally, there is currently no API access, pricing, or latency data for MiMo-V2.6, so it is too early to make product conclusions.

Sources

Author

Look at AI, editorial team