Meituan has announced the successful training of a 1.6 trillion parameter AI model, LongCat-2.0, using only domestically produced hardware. The model was trained on a cluster of over 50,000 AI ASICs and processed more than 35 trillion tokens. According to the company, this achievement demonstrates China's growing capability to develop large-scale AI models without reliance on foreign technology. The project, led by Meituan's LongCat team, has only been in existence since 2023, with its first model released late last year. The company said, "LongCat-2.0 has demonstrated that we now have the capability to train large-scale models on domestic computing clusters."
On several benchmarks, LongCat-2.0 outperforms leading Western models. It achieved scores of 59.5 on SWE-bench Pro and 77.3 on SWE-bench Multilingual, surpassing Gemini 3.1 Pro and GPT-5.5. However, it lagged behind Claude Opus 4.7 and 4.8. On other tests like IFEval, IMO-AnswerBench, and GPQA-diamond, it performed significantly lower than Gemini and GPT-5.5. The model is not yet available on HuggingFace, making independent verification challenging.
Meituan did not disclose the specific chip manufacturer used for the training. The development highlights China's progress in AI despite U.S. export controls that have been in place since 2022. The company's success in training a trillion-parameter model without foreign hardware has drawn attention to the country's growing AI capabilities.
Source: thedecoder
