Meituan Trains a 1.6T Open Model on Chinese Chips

Meituan, the Chinese delivery giant, has open-sourced a coding model that quietly climbed to the top of OpenRouter before anyone knew who built it. LongCat-2.0 is a 1.6-trillion-parameter mixture-of-experts model with a one-million-token context window, released on June 30 under a permissive MIT licence, as VentureBeat reported.

The headline is not really the size. It is the silicon. Meituan says the model was trained and served on a cluster of more than 50,000 Chinese-made ASICs, with no Nvidia GPUs involved at any stage. If that holds up, LongCat-2.0 is the first trillion-parameter model built end to end on domestic Chinese compute, at a moment when US export controls have made high-end Nvidia hardware hard to get inside China.

The stealth run

For about two months the model ran anonymously on OpenRouter under the codename Owl Alpha. It became one of the platform's most-used models, handling roughly 10 trillion tokens a month, before Meituan revealed that Owl Alpha and LongCat-2.0 were the same thing. Developers had been leaning on it heavily without knowing its origin, which is a reasonable real-world signal that the model is genuinely useful rather than just a benchmark exercise.

Read the benchmarks with care

Meituan puts LongCat-2.0 at 59.5 on SWE-Bench Pro, a shade above the 58.6 it cites for GPT-5.5. That is a margin of under a single point, and the numbers are self-reported. Until an independent group runs the tests, the fair reading is that LongCat is competitive with frontier coding models, not that it beats them. Self-reported wins have a habit of shrinking under outside scrutiny.

Even with that caveat, the release matters. It lands days after Z.ai put its own GLM-5.2 weights out to court coders, and it fits a wider pattern of Chinese labs choosing open weights as their route to influence. Giving the model away builds a developer base that no amount of marketing can buy, and it sidesteps the question of whether anyone would pay for a closed Chinese frontier model in the first place.

The domestic-hardware claim is the part worth watching. For two years the working assumption has been that serious model training needs Nvidia, and that export limits would keep China a step behind. A trillion-parameter model trained on home-grown ASICs, if the claim survives scrutiny, chips away at that assumption. It does not erase the gap, but it narrows the story the controls were meant to tell.