LongCat-2.0

For most of the spring, the most heavily used model on OpenRouter's global charts was a stealth entry called Owl Alpha, routed to by tens of thousands of developers who had no idea who built it or what it ran on. It was Meituan's — the Chinese food-delivery giant — and last week the company dropped the disguise, put the weights on Hugging Face under a permissive MIT license, and made a claim that is the actual news: LongCat-2.0, a 1.6-trillion-parameter model, was pre-trained and is served entirely on roughly 50,000 Chinese-made accelerators, with no Nvidia GPUs anywhere in the loop.

If that holds, it is the first trillion-parameter model built end to end on domestic Chinese silicon — a concrete data point in the export-controls debate that has run on speculation. The argument for cutting off Nvidia has always assumed frontier training needs it; Meituan is claiming a near-frontier coding model, on par with the strongest Western systems by its own tests, without a single one. That is what would matter, and it is also exactly the part no outsider can check. Meituan won't name the chipmaker (reporters infer Huawei), and won't disclose the cost, the training time, or the data. The benchmark scores are self-reported.

So take the chips claim as Meituan's, not as fact — it may be a stretch dressed as a milestone. But the weights are real and downloadable today, and the adoption already happened in the open: whatever it ran on, a lot of developers were quietly relying on a Chinese frontier model before anyone told them so.