Xiaomi's MiMo v2.6: When a Phone Company Started Training AI in Public

When you think of frontier AI labs, you think of OpenAI, Anthropic, Google DeepMind — names whispered with reverence and surrounded by nondisclosure agreements. You probably don't think of Xiaomi. You should.

Last week, Xiaomi's MiMo team did something no major AI lab has done before: they turned on a live dashboard showing their reinforcement learning training run in real time. Every token processed, every dollar spent, every benchmark point gained — all visible to anyone with a web browser. And the internet is watching.

The Most Expensive Livestream on Earth

The numbers are staggering. MiMo-V2.6-Pro is burning through approximately $432,000 per day — that's $5 every second — in compute costs alone. As of publication, the Pro model has processed 32.5 billion tokens, while a concurrent Flash variant has chewed through 49.4 billion tokens at an additional cost of $512,000.

This isn't a marketing stunt. It's a 1 trillion parameter-class model undergoing reinforcement learning at scale, with 2 billion tokens per step across 1,568 prompts and 16 asynchronous rollouts. The technical scope is on par with anything happening at US frontier labs — the difference is that you can watch it happen.

A 47-Point Benchmark Jump, Live

Here's where it gets interesting. The MiMo-V2.6-Pro model has already reached 65.97% on the DeepSWE v1.1 benchmark — a 47-point jump from the MiMo-V2.5 baseline of 19%. That's not a marginal improvement. That's a generational leap, happening in real time on a public dashboard.

For context, here's where the current frontier sits on the same benchmark:

• GPT-6 Astra: 74%
• Claude Fable 5: 70%
• Kimi K3: 69%
• Grok 4.6: 67%
• MiMo-V2.6-Pro: 65.97% (and still training)

Xiaomi — a company best known for phones and rice cookers — is running neck and neck with the most well-funded AI labs on the planet. And they're doing it in full view.

The Woman Behind the Run

The MiMo team is led by Luo Fuli, who previously worked at DeepSeek before joining Xiaomi. On September 16, she broke a six-month silence with a simple message: "Nearly half a year of silence. We spent it studying one problem: how far RL can scale. MiMo-V2.6 is in the middle of its RL run right now."

That's not how AI labs typically announce things. There's no press tour, no teaser campaign, no exclusive with a friendly journalist. Just a live dashboard and a tweet. It's the kind of move that makes you wonder what the rest of the industry is hiding behind their closed doors.

Transparency as Strategy

Xiaomi's approach aligns with a broader 2026 trend among Chinese AI labs. Kimi K3, Qwen 3.8, and now MiMo are all pushing toward compute and training transparency. The implicit message: we have nothing to hide because we're doing real work.

This stands in direct contrast to the Amodei pacing framework, which argues that safety requires strict coordination and controlled information flow. US labs often cite safety as justification for secrecy. Xiaomi's dashboard suggests an alternative: accountability through visibility.

There's a competitive angle too. When you're a phone company trying to establish credibility in AI, showing your work is more valuable than claiming results. Anyone can put up a benchmark number. Not anyone can show you the training run that produced it.

The Catch

It's not all sunshine and open dashboards. The Hacker News community spotted some issues. Dashboard numbers may reset or replay upon page refresh, raising questions about the authenticity of the real-time data. More notably, observers found a line item labeled "Claude Distill Requests: hidden" on the dashboard.

That's a problem. It suggests MiMo may still rely on distillation from Anthropic's Claude models — meaning the training isn't fully independent. It's one thing to show your RL run. It's another to show the proprietary model you're distilling from. The word "hidden" does a lot of heavy lifting there.

Still, even with that caveat, the level of transparency here exceeds anything from US frontier labs. You can argue about what's hidden, but you can't argue that the visible part is more than anyone else is showing.

What This Means for the Industry

The MiMo v2.6 run is a test case for whether public, high-cost training can coexist with competitive pressure. If Xiaomi produces a frontier-grade model while showing their work, the argument for closed-door development gets significantly weaker.

It also forces a conversation about what "safety" really means in AI development. Is it about controlling information? Or is it about making the process visible enough that problems can be caught by the community? The MiMo team seems to believe the latter.

Luo Fuli has stated that the team intends to open-source the technical details incrementally, following the precedent set by the MIT-licensed MiMo-V2.5. If they follow through, the full recipe for a frontier-grade RL training run will be public knowledge.

The Bottom Line

A phone company is spending $432,000 a day to train a trillion-parameter AI model in full public view, and it's working. The model is keeping pace with the best systems from OpenAI, Anthropic, and Google. The dashboard is live. The benchmarks are real. And the rest of the industry is watching through a one-way mirror — they can see out, but we can't see in.

Whether this is a genuine shift in research culture or a strategic performance, Xiaomi has changed the conversation. The question now is whether US labs will respond with their own transparency — or retreat further behind closed doors.

One thing's for certain: the next time you think about who's pushing the frontier of AI, remember that a company that makes rice cookers is doing it live, with the receipts.