China Might Be Leading the AI Race

In our last post, we looked at how the US government is moving to restrict Chinese AI models. The bigger story is why they feel the need to: Chinese labs are no longer catching up. They are shipping frontier models.

The clearest proof came in July. Moonshot AI released Kimi K3, an open-weight model with 2.8 trillion parameters and a 1 million token context window. Within hours it took the #1 spot on Arena's Frontend Code leaderboard. That put it ahead of Claude Fable 5, the current leading coding model, 1,679 points to 1,631, decided by 1,757 blind developer votes. It won six of the seven frontend categories. No open-weight model had ever led that board before.

Kimi K3 tops the Frontend Code Arena with 1,679 points, ahead of Claude Fable 5 at 1,631.

To be fair, Moonshot's own technical blog admits K3 still trails Claude Fable 5 and GPT 5.6 Sol on most overall benchmarks. But that is exactly the point. An open model you can download is now trading wins with the best closed models in the world.

The market noticed. Demand surged sixfold after launch, and Reuters reported that Moonshot had to pause new subscriptions because it ran out of compute. The company is now pushing toward a Hong Kong IPO at a reported $30 billion valuation.

The full open weights were released on July 27, 2026 on the Moonshot AI Hugging Face page. In theory you can run K3 locally. In practice a 2.8T parameter model needs a beast of a machine, so most people will use it through a hosted API instead.

## Not Just Code: Images and Video

Coding is not the only front. Alibaba's Qwen-Image-3.0-Pro now sits at #5 on the Text-to-Image Arena with 1,263 points, ahead of Nano Banana Pro and level with Nano Banana 2. That is a 72-point jump over the previous generation.

Qwen-Image-3.0-Pro at #5 on the Text-to-Image Arena with 1,263 points.

One caveat: Alibaba has not released the 3.0 weights yet. But the previous generation, Qwen-Image-2512, is fully open, and it already competes with Google's best.

Video is where China is clearly in front. ByteDance's Seedance 2.5 generates native 30-second clips in a single pass, and Kuaishou's Kling 3.0 reached $300 million in annual revenue faster than any AI video platform in history.

Neither is open-weight, but the pricing changes what is possible. A 30-second product video now costs a few dollars to generate, against the hundreds or thousands a production agency would charge. Brands are already running real ads made this way.

The video below was generated entirely with Seedance 2.5.

Why America Is Worried

The US concern comes down to two things: security and market share.

The security worry is about data. If US companies build on Chinese-hosted AI services, could user data or internal systems be exposed or exploited? For APIs served from Chinese infrastructure, that is a fair question to ask.

Dario Amodei, Anthropic's CEO, has been one of the loudest voices. He has accused Chinese labs of training their models by distilling the outputs of top US frontier models. Critics push back that some of this alarm is less about safety and more about shutting out cheaper competition.

The market share worry is simpler. K3's API costs $3 per million input tokens and $15 per million output, and cache hits drop input to $0.30. Claude Fable 5 costs $10 in and $50 out. When an open model is a third of the price and wins leaderboards, the pressure on US pricing is real.

What You Can Do With This

This is the part most coverage skips. Here is what these models mean for you today.

If you build software. Try Kimi K3 on front-end work. It leads the arena there, at roughly a third of Fable 5's price per token. You can reach it through Moonshot's API or a router like OpenRouter without changing your workflow.

If you run a business. The Seedance video above is the takeaway. Upload your product photos, describe the scene, and you get a polished ad video for a few dollars instead of a videographer's invoice. For Reels and product ads, this is usable today.

If security is your blocker. Here is the move most people miss. Because K3 is open-weight, you do not have to send anything to a Chinese server. US providers like Fireworks and Together AI host the same model on American infrastructure, some with zero data retention options. You keep the price and the quality, and the data question mostly disappears.
That last option is what makes open weights different from a cheap API. The model is an artifact you can inspect and run wherever you choose. A closed model can never give you that.

The Frontier Has Two Flags Now

China is not trying to catch up anymore. With Kimi K3 on top of the frontend arena, Qwen in the image top five, and Seedance and Kling leading video, Chinese labs are part of the frontier.

For builders, the practical question is not whose flag is on the model. It is what you can ship with it, at what price, and on whose servers. Right now, some of the best answers to that question are coming out of China.

---

Originally published at domlabs.dev.