On Monday, Xiaomi put the world's top-ranked open-weight AI model on Hugging Face. Free to download. MIT license. Available right now.
MiMo-V2.6-Pro scores 46 on Artificial Analysis's Intelligence Index — the highest any open-weight model has ever hit. It runs within two points of closed frontier models on software engineering benchmarks, and it actually outperforms GPT-5.6 Sol on Terminal Bench (89.9%). The Flash variant is smaller, cheaper to deploy, and competitive with models that cost 20x more per token via API.
This isn't "good for an open-source model." It's just good.
Now here's the part that puts a different frame on it.
Two weeks ago, CISA, the FBI, and NSA issued a joint advisory accusing six Chinese AI companies of industrial-scale distillation attacks against US AI providers. Three days later, a 60-page threat intelligence report named seven labs by name: Alibaba, Moonshot AI, DeepSeek, Zhipu, Xiaomi, MiniMax, and SenseTime.
The combined finding: roughly 200 million exchanges extracted from at least one closed-model provider's API over about eight months. Alibaba's campaign was the biggest — 151 million exchanges between May and July 2026.
Xiaomi's specific case (GTG-16008): 400,000 exchanges routed through 1,500 proxy accounts over a 20-day window in March and April. The technique involved coding harnesses called OpenClaw and OpenCode, which replayed users' actual conversations from Xiaomi's own apps through a closed API. Those outputs got labelled as training data.
The idea is simple enough that it's almost elegant. You can't download a closed frontier model. But you can query it at scale. Route enough real user conversations through the API, capture the outputs, and you've bought yourself a distilled copy of the model's capabilities at a fraction of the compute cost of training from scratch. The threat intelligence report estimated Xiaomi spent about $3.47 million on compute to train MiMo-V2.6. The model it was allegedly distilling from costs far more to replicate independently.
Xiaomi has not responded to the allegations. The model was released one day before the first-ever UN Security Council session to seat US and Chinese frontier AI developers in the same room.
What you should actually do with this
The MIT license is clean. You are not exposed as a downstream user. MIT is MIT — whatever happened between labs and their API providers is a terms-of-service dispute between those parties, not a liability that flows to you. That's worth saying plainly because the government advisory will make some procurement teams nervous.
That said, if your organization has formal AI procurement policies, this week is a reasonable time to document your evaluation: you knew about the advisory, you assessed the MIT license status, here's your reasoning. That takes 20 minutes and it's a better position than discovering the question later.
On the practical side, the model is worth testing.
MiMo-V2.6-Flash is available via Xiaomi's API at $0.04 per million input tokens, $0.16 per million output. For comparison, GPT-5.6-class access runs around $4/$16. If you have a workflow burning $200/month on AI API costs, this could be $8. Run your actual workload through both endpoints and see what the quality difference is. For most summarization, classification, or coding-assist tasks, Flash will be close enough.
If you need private inference — healthcare, legal, government, anything with data residency rules — the Flash weights are downloadable today. VentureBeat's headline was "Better than DeepSeek." For self-hosted private AI, that comparison matters. DeepSeek R1 has been the default recommendation for on-premise deployment for most of 2026; MiMo-V2.6-Flash is the new benchmark to beat.
The signal under the noise
Something worth understanding: the open-weight/closed-model capability gap has been closing faster than most researchers expected. The common explanation has been better training techniques, cheaper compute, and improved data curation. The picture emerging from these advisories is that a chunk of the gap closed because of distillation — labs replicating proprietary capabilities by routing real user workloads through closed APIs at scale.
This cuts in two directions. The good one: you now have access to something your team can run, own, and modify without paying ongoing API costs to a vendor who can reprice or shut off access at any time. MIT-licensed frontier-class AI is not something that existed six months ago.
The uncomfortable one: if distillation at this scale continues, the incentive to do expensive, novel research and publish it openly gets smaller. The organizations that built the models being scraped are now less likely to release the next round publicly. You can already see this in how model releases have shifted — fewer open weights, more API-only, more restrictions on commercial use.
What that means for your team: the window where MIT-licensed frontier AI is available may be shorter than it looks. MiMo-V2.6 is real, the license is real, and the API pricing is real right now.
Benchmarking an open-weight model against your actual workload, setting up private inference, and figuring out when it makes sense to cut over from a paid API — that's the kind of sprint we run with teams. Two weeks, working system at the end. If you want a second opinion on whether Flash is ready for your use case, we're easy to reach.
Sources: VentureBeat · The Hacker News on the distillation report · Xiaomi official release notes · Hugging Face weights