Meituan China trains a 1.6T param model without Nvidia!

Food delivery platform Meituan (but really a tech company), has released LongCat-2.0, a language model with 1.6 trillion parameters. According to Meituan, the model shows that cutting-edge AI can be trained on domestic computing platforms.

1No Nvidia: What's Under the Hood

The model was trained on a cluster of over 50,000 domestic AI accelerators without any Nvidia chips. The cluster handled more than 35 trillion tokens. The large-model team was established in 2023 and released their first model in late 2025.

2Benchmarks: Where It Leads, Where It Lags

On agentic coding benchmarks like SWE-bench Pro (59.5) and SWE-bench Multilingual (77.3), LongCat-2.0 performs better than some Western models. But on other benchmarks like IFEval (90.0), IMO-AnswerBench (81.8) and GPQA-diamond (88.9), it lags behind top Western models. Also, note that the model has not been published on HuggingFace yet, and we don’t know which chipset manufacturer it uses. So take the reported scores with a grain of salt.

Assuming the scores are true, this would be a powerful rebuttal to the argument that U.S. restrictions on exporting Nvidia chips (since 2022) effectively prevent China from advancing in cutting-edge AI. Of course, this isn’t the only instance of Chinese companies making this claim. But it’s one of the clearest examples with quantitative data on hardware capacity and training data size.