LongCat-2.0: China Just Open-Sourced a Frontier AI Built With No American Chips

Getting your Trinity Audio player ready...

LongCat-2.0 is the kind of announcement that sounds technical and lands geopolitical. On 30 June, Chinese delivery giant Meituan open-sourced a 1.6-trillion-parameter AI model — and the detail that matters is not the size. It is that the entire thing, from training to inference, ran on Chinese chips.

For years, Washington’s strategy for containing China’s AI rise rested on one chokepoint: cutting off access to Nvidia’s top processors. LongCat-2.0 was trained on a cluster of more than 50,000 domestic Chinese ASICs. No Nvidia. No loopholes. And the result is not a distant also-ran — before Meituan revealed its identity, the model had been quietly topping usage charts on OpenRouter under the codename “Owl Alpha,” with agentic-coding scores placing it in the same conversation as GPT-5.5, Gemini 3.1 Pro and Claude Opus 4.6.

Then Meituan did the most disruptive thing possible: it gave the model away, under a permissive MIT license that lets any company anywhere use it commercially, for free.

What makes LongCat-2.0 technically remarkable

The engineering deserves respect on its own terms. LongCat-2.0 is a mixture-of-experts system: although it holds 1.6 trillion parameters, a token-level routing scheme activates only about 48 billion of them per token, which keeps running costs far below what the headline number suggests. It ships with a native one-million-token context window — enough to hold entire codebases or document archives in a single prompt.

That a food-delivery company built this is less strange than it sounds. Meituan runs one of the world’s most complex logistics operations and has been pouring its profits into AI. But the symbolism is hard to miss: the second tier of China’s tech industry is now producing near-frontier models, not just its giants.

The uncomfortable question for Washington

Honesty requires caution here. “Trained on domestic chips” is Meituan’s claim, and independent verification of training details is notoriously difficult. Sceptics will reasonably ask about chip provenance and benchmark selection. And near-frontier is not frontier — the very best American models retain an edge.

But the direction of travel is unmistakable, and it poses a question export controls were never designed to answer. The controls assumed China could not scale AI without American silicon. LongCat-2.0 is a 1.6-trillion-parameter argument to the contrary. Worse for the containment strategy, the MIT license turns the model into infrastructure for the whole world: every startup in Jakarta, Lagos or Dhaka that builds on a free Chinese model is a small shift in where the AI ecosystem’s centre of gravity sits.

Chokepoints work until the other side builds around them. This week suggests that moment is no longer hypothetical — and the race, from here, gets decided by invention rather than embargo.

---

Fact basis: Meituan open-sourced LongCat-2.0 on 30 June 2026; 1.6T-parameter MoE, ~48B active per token; 1M-token context; MIT license; trained/inferenced on 50,000+ domestic Chinese ASICs (company claim); previously topped OpenRouter as “Owl Alpha”; benchmark comparisons vs GPT-5.5, Gemini 3.1 Pro, Claude Opus 4.6 (VentureBeat, SCMP, Reuters).