China trained a frontier AI model without a single Nvidia chip
Meituan's LongCat-2.0 proves that US chip export controls are not enough to keep China from building world-class AI.
- 1.6T
- total model parameters
- 50,000+
- Chinese-made chips used for training
- 59.5
- coding test score, edging out GPT-5.5 (58.6)
- $0.30
- API price per million tokens
For four years, the United States has blocked China from buying Nvidia's most powerful AI chips. The idea was simple: no advanced chips, no frontier AI. That assumption just broke. On June 30, Chinese tech company Meituan released LongCat-2.0 — a massive AI model trained entirely on Chinese-made hardware, with zero Nvidia chips involved.
LongCat-2.0 has 1.6 trillion parameters — think of parameters as the internal settings that shape how an AI thinks. Meituan says it scored 59.5 on SWE-bench Pro, a standard coding test, edging out GPT-5.5's 58.6. Training ran on more than 50,000 Huawei-made chips across 35 trillion tokens of data. China now has a frontier AI model built on hardware it makes itself.
What makes this story more interesting is how Meituan launched it. For roughly two months before the announcement, the model ran quietly on OpenRouter — a platform where developers try different AI tools — under a fake name: "Owl Alpha." Developers used it heavily and rated it highly without knowing who built it. The real-world test came before the press release. That is an unusual kind of confidence.
The model is open source with an MIT license — anyone can download it, use it, and build commercial products with it, no restrictions. API pricing starts at $0.30 per million tokens, well below comparable US models. A powerful, cheap coding AI with no access restrictions just entered the market. For any team that uses AI to write software, it belongs on your evaluation list.
The bigger signal is for the whole industry. The chip export controls were supposed to slow China's AI progress. LongCat-2.0 shows that bet is working less well than expected. The AI landscape is splitting — more models, more sources, more competition. That is mostly good news for builders: more choice, lower prices, and fewer single points of dependency.