Meituan launches LongCat‑2.0, a trillion‑parameter AI model trained on domestic Chinese chips

Chinese technology firm Meituan announced the release of LongCat‑2.0, a large‑scale language model with roughly a trillion‑plus parameters. The model was trained and can be run entirely on domestically produced Chinese AI chips, using a 50,000‑card compute cluster that does not rely on Nvidia GPUs subject to U.S. export controls.

LongCat‑2.0 is open‑sourced under an MIT licence and made available through Hugging Face and the OpenRouter API. Benchmarks cite performance comparable to Google’s Gemini 3.1, with scores such as 59.5 on the SWE‑bench Pro coding test. The launch underscores China’s push for AI hardware independence amid ongoing U.S. semiconductor restrictions and positions Meituan against other Chinese AI developers like ByteDance’s Doubao and DeepSeek.