LongCat-2.0: China’s 1.6T Model Trained Without Nvidia
3 min readChina just cleared a milestone that US export controls were designed to prevent. On June 30, Chinese services giant Meituan open-sourced LongCat-2.0, a 1.6 trillion parameter model that the company says was trained and served end to end on domestic Chinese chips, with no Nvidia GPUs involved. It is the largest AI model yet claimed to be built entirely on local silicon.
Why Domestic Chips Matter
For two years, US restrictions have limited China’s access to Nvidia’s most powerful accelerators, the hardware that trains frontier models. Chinese labs adapted, but with a catch. Earlier flagships like DeepSeek leaned on domestic chips for inference, the everyday job of answering prompts, while still relying on foreign silicon for the compute heavy pretraining phase. As SiliconANGLE reports, LongCat-2.0 is notable because Meituan says it removed that dependency entirely.
What LongCat-2.0 Actually Is
LongCat-2.0 is a Mixture of Experts model with 1.6 trillion total parameters, though it activates only about 48 billion per token to keep costs down. It ships with a native context window of one million tokens under a permissive MIT license. Meituan says it trained the model on a cluster of more than 50,000 domestic ASICs, custom chips built specifically for AI work.
The model is not a newcomer. According to VentureBeat, it quietly topped the OpenRouter developer charts for two months under the codename “Owl Alpha” before Meituan revealed what it was. On agentic coding benchmarks, its scores land near GPT-5.5, Gemini 3.1 Pro, and Claude Opus 4.6. One caveat holds it back for now: the full weights are not posted yet, with the GitHub and Hugging Face pages still showing a “coming soon” note.
Why It Matters
The claim, if it holds up, undercuts a core assumption behind US chip policy: that cutting off Nvidia access would keep China a generation behind at the frontier. A near frontier model trained without a single Nvidia GPU suggests domestic alternatives have matured faster than many expected. Releasing it under an open license also adds fuel to a strategy Chinese firms have leaned into all year, giving away capable models to win global developer share while several US labs move to gate access.
The next thing to watch is verification. Independent researchers will want to confirm the training claims once the weights land, and rivals will study whether a cluster of domestic ASICs can really match a Western GPU farm at scale.
For now, LongCat-2.0 stands as a signal that the hardware gap Washington hoped to widen may be narrowing instead.