Xiaomi Live-Streamed Its RL Training. The Bill Was the Point.

On September 22, Xiaomi released MiMo-V2.6, an open-weights model family with two variants: Pro and Flash. Both are natively multimodal, support a 1M-token context window, and ship under an MIT license.

Pro scored 46 on the Artificial Analysis Intelligence Index — the highest of any open-weights model, ahead of Kimi K3 at 44 and GLM-5.3 at 45. Xiaomi acknowledged in its own materials that Claude Fable 5.1 and GPT-6 Astra both sit at 53.

Model specifications

The leaderboard placement is the headline. The training dashboard is the story.

The Training Was Public. So Was the Bill.

MiMo team lead Luo Fuli opened a live training page before the launch. It reads directly from the trainer logs and shows step count, token consumption, reward curves, cost accumulation, and hardware events.

The page showed Pro consuming roughly $20,000 per hour and Flash roughly $10,000 per hour — over $30,000 combined. By September 18, the two runs had logged 11 restarts, 9 of them on the Pro line. GPU memory faults and recovery events were visible in the log.

Training cost, final

Publishing this data is unusual at the frontier. Training costs are typically disclosed selectively, if at all, and never in real time. Xiaomi turned the expense into an auditable record. The 46 AA score was produced by Artificial Analysis using its own framework, not reported by Xiaomi — which is the detail that makes the comparison meaningful.

Flash Is the Model You'll Actually Deploy

Pro gets the attention because it tops the open-weights list. Flash may matter more in production.

On CyberGym, Flash scores 95.1 — above Pro's 94.0. The pricing gap is wider than the performance gap.

API pricing (per million tokens)

Xiaomi says pricing is unchanged from V2.5 — "smarter at the same price." The company claims that at equivalent intelligence levels, its models cost 1/20 to 1/60 of overseas frontier models.

For high-frequency agent workflows making thousands of calls a day, the three-fold gap between Pro and Flash determines whether a deployment is viable at all. Both models share the same 1M-token context window and multimodal input support.

Agent benchmark comparison

The 72.57 on DeepSWE Is Not a Leaderboard Score

Xiaomi's dashboard reports DeepSWE v1.1 improving from 58.4 to 72.57 during RL training. The number looks close to the public leaderboard's top score of 74.0 held by Claude Opus 5.

It is not comparable.

The 72.57 comes from Xiaomi's own harness, its own grader, and its own offline run. It was not submitted to the public DeepSWE leaderboard, and it has not been independently reproduced. Public leaderboard scores include a submission configuration, Pass@1, confidence intervals, and an average task cost. Xiaomi's internal number includes none of those.

This is not an accusation of dishonesty. A vendor harness can be entirely honest and still not be a leaderboard score. The distinction matters because the two numbers will be read side by side, and only one of them is an independent measurement.

The Artificial Analysis 46 is the comparable figure. The DeepSWE 72.57 is a training-progress metric.

What the Release Actually Establishes

Three things are verifiable. The AA score of 46 is third-party and reproducible. The training cost is a public record with timestamps and restart events. The weights are open under MIT.

Two things are not yet established. Whether MiMo-V2.6-Pro is deployable at scale — Artificial Analysis currently tracks only one API provider, compared to 15 for MiniMax M3. And whether the Flash-optimized configuration holds up in production, where agent workloads differ from benchmark conditions.

The live training dashboard was a deliberate choice. It answers a question the industry usually leaves opaque — what does it cost to train a frontier-adjacent open model, and how often does the hardware fail along the way? Xiaomi published the answer. The model is the product. The bill is the proof.

P.S. MiMo V2.6-Pro is currently available from a single API provider on Artificial Analysis, while MiniMax M3 lists 15. Open weights do not guarantee broad serving. The model can be downloaded and run, but for teams that need a hosted endpoint with redundancy, availability is still the constraint.

Frequently Asked Questions

Q: What did Xiaomi release?

A: MiMo-V2.6, an open-weights model family with two variants: Pro (1.02T total parameters, 42B active) and Flash (309B total, 15B active). Both are multimodal with 1M-token context, under an MIT license.

Q: What is UltraSpeed?

A: It is not a separate model. It is a service tier for Pro that runs the same weights at higher speed, priced about 10 times standard Pro.

Q: What was the live training dashboard?

A: MiMo team lead Luo Fuli opened a public training page that read directly from trainer logs, showing cost accumulation, restarts, and GPU faults in real time. Pro cost about $2.62 million; Flash about $850,000.

Q: Why does Flash matter more than Pro?

A: Flash trails Pro by narrow margins on most agent benchmarks but costs one-third as much per output token. On CyberGym it scores 95.1, above Pro's 94.0. For high-frequency agent deployments, it is the practical choice.

Q: Is the DeepSWE 72.57 a leaderboard score?

A: No. It comes from Xiaomi's own harness and grader, not a public submission. Public leaderboard scores include Pass@1, confidence intervals, and average task cost. The comparable figure is the AA score of 46.