News
Meituan Open-Sources LongCat-2.0, the Secret OpenRouter Hit Trained Without Nvidia
Chinese food delivery giant Meituan has unveiled LongCat-2.0, a 1.6-trillion-parameter model trained on more than 50,000 domestic chips without a single Nvidia unit. The model anonymously topped OpenRouter's rankings for two months under the codename Owl Alpha before Meituan revealed itself as its creator.
Meituan, the company best known in China for food delivery, has open-sourced LongCat-2.0, a 1.6-trillion-parameter system that anonymously dominated popularity rankings on the OpenRouter platform for two months before anyone learned who was behind it.
The model landed on Hugging Face and GitHub as a mixture-of-experts architecture with a native 1-million-token context window, built specifically for agentic coding. At launch the model weights themselves were marked "coming soon," which alone sparked debate in the open source community over whether this counts as a genuinely open release.
Two months in hiding
The most surprising part of the story is that LongCat-2.0 didn't come out of nowhere. For roughly two months it ran on OpenRouter under the anonymous codename Owl Alpha, generating more than 10.1 trillion tokens per month, with a daily average reaching 559 billion and usage growing 242 percent month over month. Only after climbing to the top of the platform's rankings did Meituan reveal that the model was theirs.
This approach, a quiet launch under a false name, testing market reaction, and only then an official reveal, is becoming increasingly popular among large language model developers, since it lets them gauge genuine user interest without the marketing pressure that comes with a high-profile launch.
Chinese hardware, no US chips
What matters most about this release is how the model was built. LongCat-2.0 was trained from scratch on a cluster of more than 50,000 domestically designed ASIC chips made in China, without a single Nvidia GPU involved. According to its creators, that makes it the first trillion-parameter-class model trained and run end to end on computing hardware independent of the United States.
Given US export restrictions on advanced AI chips to China, this sends a strong signal that Chinese companies can build frontier-class models without access to Nvidia's best chips. Over its full training and validation cycle, the model consumed more than 35 trillion tokens of data.
Results and pricing rival the West
On benchmark tests, LongCat-2.0 posts results comparable to leading Western models: 59.5 points on SWE-bench Pro (narrowly ahead of GPT-5.5's 58.6), 70.8 points on Terminal-Bench 2.1, which measures agentic terminal tasks, and 79.9 points on BrowseComp, a web-browsing test. Promotional starting prices are $0.30 per million input tokens and $1.20 per million output tokens, well below Anthropic's or OpenAI's rates.
For Polish companies and developers, this means yet another cheap alternative to Western-provider models, joining already-available open Chinese models like GLM-5.2. The growing number of competitively priced Chinese models is putting pressure on leading providers to cut prices and is widening the choice for companies building their own agentic tools.
The LongCat-2.0 launch fits into a broader trend of recent months, in which Chinese tech giants, not just specialized AI labs but also companies from entirely different industries like food delivery, are pouring billions into building their own computing infrastructure and language models, treating it as part of the country's strategic drive for technological independence.
Sources: VentureBeat (venturebeat.com), Geopolitechs (geopolitechs.org), Digital Today (digitaltoday.co.kr)
