China just trained a 1.6T parameter model without Nvidia
meituan trained the 1.6 trillion parameter longcat-2.0 entirely on domestic chips, showing china can bypass us export controls to build competitive frontier models.
meituan has trained longcat-2.0, a 1.6 trillion parameter ai model, using a cluster of over 50,000 domestically produced asics, processing more than 35 trillion tokens. benchmark results show it outperforming western models like gemini 3.1 pro and gpt-5.5 on swe-bench pro and multilingual coding tasks, though it lags behind claude on reasoning tests. this demonstrates that chinese firms can effectively bypass us semiconductor restrictions to build competitive infrastructure. the specific chip manufacturer remains undisclosed, and the model has not been released for independent verification.