Meituan open-sources LongCat-2.0: 1.6T model, trained without Nvidia

In late June 2026 Meituan — China's largest food-delivery and local-services platform — open-sourced LongCat-2.0 under a permissive MIT licence. It's a 1.6-trillion-parameter Mixture-of-Experts model with a native 1-million-token context, activating roughly 33-56 billion parameters per token. Meituan also revealed it was the model behind 'Owl Alpha', the anonymous stealth entry that had topped OpenRouter's developer usage charts for about two months. The headline claim: it's the first trillion-parameter model trained and served end-to-end on Chinese-made chips — a cluster of more than 50,000 domestic ASICs, no Nvidia GPUs.

Read the wins carefully. The benchmarks are Meituan-reported and not independently verified: 59.5 on SWE-Bench Pro (vs GPT-5.5's 58.6), 70.8 on Terminal-Bench 2.1, 77.3 on SWE-Bench Multilingual. The much-repeated 'beats GPT-5.5' line holds on exactly one test by under a point — a statistical tie, not a settled ranking. At announcement Meituan also listed the full weights as 'coming soon' rather than all posted, and the non-Nvidia training claim is the company's own until someone reproduces it.

Watch three things: whether the full weights ship as promised, whether independent evaluators confirm the coding scores, and whether anyone outside Meituan can reproduce the no-Nvidia training claim. The wider signal is already clear — Chinese open-weight models now make up roughly 61% of OpenRouter's top-10 traffic, and LongCat-2.0 is the heaviest entry yet in that surge.