Meituan, the Beijing food-delivery company often described abroad as China’s DoorDash, open-sourced a 1.6-trillion-parameter model on Tuesday and said the entire training run happened on Chinese silicon. The company is calling LongCat-2.0 “the industry’s first trillion-parameter model to complete full-process training and inference on a 50,000-card domestic computing power cluster.” It declined to name the chipmaker.

That last detail is the entire story. Export controls were designed to make a sentence like Meituan’s structurally impossible, and the rhetorical posture of the announcement, MIT-licensed weights, a 1-million-token context window, claimed parity with Google’s Gemini, OpenAI’s GPT-5.5, and Anthropic’s Claude Opus on coding and agent benchmarks, is calibrated to advertise that impossibility as resolved. Reuters notes the benchmark claims are Meituan’s, not independently verified. The Next Web flags that the hardware claim itself rests on Meituan’s account of its own infrastructure.

The strategic backdrop matters. Reuters reports that domestic chipmakers including Huawei and Enflame have been moving quickly into the vacuum left by U.S. suppliers, signing AI developers through supply deals. A LongCat-sized training run is the kind of marquee customer those vendors need to legitimize the stack.

It’s also not a side project. SiliconAngle’s Kyt Dotson notes Meituan “jumped into the AI model development scene as far back as 2023 when it acquired the startup Light Year Beyond for $281 million,” with its own model effort announced in 2025. VentureBeat’s Carl Franzen reports Meituan also “unmasked the model as the computational engine behind ‘Owl Alpha,’ the anonymous stealth model that has spent the last two months commanding global developer charts on OpenRouter,” now running roughly 10.1 trillion tokens a month, averaging 559 billion per day, up 242% month-over-month. The architecture activates around 48 billion parameters per token.

One caveat sits on the repos themselves: both still read “coming soon.” The weights aren’t fully public. In the meantime, the announcement does its job, slotting LongCat-2.0 alongside DeepSeek’s V4-pro as evidence that the sanctions regime is producing exactly the domestic capability it was designed to forestall. Washington has seen this movie before, in semiconductors after the 1980s Toshiba-Kongsberg controls, and the sequel tends to rhyme.

Sources