Pro TV asked me about Kimi K3, the newest model out of China — one that arrived free to download and immediately had the White House debating whether to restrict it.
Kimi K3 comes from Moonshot AI in Beijing, and it was shown at the World AI Conference in Shanghai this week. What makes it notable isn't only that it performs at the level of the established Western models on coding benchmarks — it's that the weights are open. You don't buy a licence, you don't call an API you don't control. You download it and you run it.
"Kimi K3 is an open-source model — open and free, in the sense that you don't have to buy it — and it reaches the level of the top, established models. These models have come very close. But this isn't new. They work: if you actually use them, they perform at the same level as the closed ones."
The part I kept coming back to on camera is that last sentence: this isn't new. The gap between open Chinese models and closed Western ones has been narrowing for well over a year, one release at a time. What's changed with K3 is the framing — it's now hard to argue that you have to pay to get frontier-level output for a large class of tasks.
Why Washington is nervous
The White House is reportedly weighing measures against Chinese open models. But there's a practical problem with a ban: open weights don't sit on a server someone can switch off. Once a model has been downloaded a few hundred thousand times, it's on laptops and in company data centres everywhere, and restricting it mostly restricts the people who would have complied anyway.
What it means in practice
For a company here, the interesting consequence isn't geopolitical. It's that a serious model can now run on your own infrastructure — your data never leaves the building, and the cost per request stops being a subscription. That's worth testing. And test it the honest way: not against a benchmark table, but against your own tasks, with your own documents. Benchmarks tell you a model is close. Only your work tells you whether close is close enough.