Can you see MiMo-V2.6's benchmark scores on the dashboard?

The MiMo-V2.6 dashboard streams metrics from the run in progress — task composition, context length and per-step timing — rather than a finished scorecard. Xiaomi had not published final model benchmarks or specifications when the livestream opened, because the reinforcement-learning phase is still going. What you are watching is the curve, not the result.

Is Xiaomi open-sourcing the MiMo-V2.6 RL recipe?

Not as one drop. Fuli Luo said the MiMo team will open-source the details piece by piece over the coming weeks instead of publishing a single comprehensive report. So the methodology behind the run is not available yet; the live dashboard is the only public window into it at the moment.

How big is each training step?

Each step of the MiMo-V2.6 run draws 1,568 prompts and generates 16 rollouts per prompt, which works out to roughly 2 billion tokens of generated text per step. Several task families — code, general reasoning, visual, cybersecurity and chat agents — are trained together in one multi-harness run rather than in separate specialist stages.

Why is a public training dashboard unusual?

Labs almost never show a run while it is happening. Reward curves and eval metrics mid-training reveal how well a recipe is working, which is competitively sensitive, so they normally appear only in a report after the fact. Xiaomi's MiMo-V2.6 livestream inverts that, and the Hacker News thread on it drew 494 points and 139 comments in under a day.

Xiaomi · 2026-09-17 · major

Xiaomi livestreams MiMo-V2.6's RL run — a public training dashboard

Xiaomi opened a live dashboard for the reinforcement-learning run behind MiMo-V2.6, streaming per-step metrics as they happen. Each step uses 1,568 prompts with 16 rollouts each, roughly 2 billion tokens.

Cover graphic for a write-up of Xiaomi's livestreamed MiMo-V2.6 reinforcement-learning run — explainx.ai

Xiaomi put the reward curves of a frontier-scale RL run on a public web page while the run is still going.

Quick facts

Maker	Xiaomi MiMo team
Model	MiMo-V2.6 — still mid-training
Dashboard	mimo.xiaomi.com/rl
Per step	1,568 prompts x 16 rollouts, about 2B tokens
Task mix	Code, general, visual, cybersecurity and chat agents
Announced by	Fuli Luo, MiMo team lead
Open source	Not yet — details promised over the coming weeks

What is it?

A live dashboard at mimo.xiaomi.com/rl now streams the reinforcement-learning post-training of MiMo-V2.6 as it happens. Fuli Luo, who leads Xiaomi's MiMo team, announced it on 17 September 2026 after what he described as nearly half a year of silence spent on one question: how far reinforcement learning actually scales. The model itself is not released — the training run is what is on display.

How does it work?

Every step of the run pulls 1,568 prompts, samples 16 rollouts from each, and works through roughly 2 billion generated tokens. Code, general, visual, cybersecurity and chat-agent tasks are trained together in a single multi-harness setup rather than one specialist run per domain, and rewards combine test-case checks with rubric grading under what the team calls agentic in-group credit assignment. The dashboard exposes task composition, context length and time per step live.

Why does it matter?

Anyone trying to scale reinforcement learning is working mostly from published post-mortems, where the failed variants have been edited out. Watching a real run tick over gives researchers a reference for what healthy curves and step costs look like at this size. The recipe is not open yet, so the dashboard is a window rather than a manual — but it is a far more honest one than a results table.

Who is it for?

RL researchers, post-training engineers, people tracking Chinese labs

Frequently asked questions

Can you see MiMo-V2.6's benchmark scores on the dashboard?: The MiMo-V2.6 dashboard streams metrics from the run in progress — task composition, context length and per-step timing — rather than a finished scorecard. Xiaomi had not published final model benchmarks or specifications when the livestream opened, because the reinforcement-learning phase is still going. What you are watching is the curve, not the result.
Is Xiaomi open-sourcing the MiMo-V2.6 RL recipe?: Not as one drop. Fuli Luo said the MiMo team will open-source the details piece by piece over the coming weeks instead of publishing a single comprehensive report. So the methodology behind the run is not available yet; the live dashboard is the only public window into it at the moment.
How big is each training step?: Each step of the MiMo-V2.6 run draws 1,568 prompts and generates 16 rollouts per prompt, which works out to roughly 2 billion tokens of generated text per step. Several task families — code, general reasoning, visual, cybersecurity and chat agents — are trained together in one multi-harness run rather than in separate specialist stages.
Why is a public training dashboard unusual?: Labs almost never show a run while it is happening. Reward curves and eval metrics mid-training reveal how well a recipe is working, which is competitively sensitive, so they normally appear only in a report after the fact. Xiaomi's MiMo-V2.6 livestream inverts that, and the Hacker News thread on it drew 494 points and 139 comments in under a day.

Try it

https://mimo.xiaomi.com/rl/