Meituan unveils 1.6‑trillion‑parameter LongCat‑2.0 model trained on Chinese chips

China’s food‑delivery giant Meituan has open‑sourced LongCat‑2.0, a large‑language model with 1.6 trillion parameters and a 1 million‑token context window. The model was trained and run entirely on a domestic compute cluster of roughly 50 000 AI ASIC “superpods”, using Huawei’s Collective Communication Library for chip‑to‑chip coordination. It is the first trillion‑parameter system fully built on Chinese‑made processors, marking a milestone in China’s push for AI self‑sufficiency amid U.S. export controls on Nvidia GPUs.

LongCat‑2.0 adopts a sparse mixture‑of‑experts architecture, activating about 48 billion parameters per token, and is positioned for agentic coding, long‑context document processing, and other complex AI‑agent tasks. Meituan claims the model matches or exceeds leading proprietary systems such as Google’s Gemini 3.1 Pro, OpenAI’s GPT‑5.5 and Anthropic’s Claude Opus on various benchmarks, including a 59.5 score on the SWE‑bench Pro test. Before its official reveal the model ranked among the top three on the OpenRouter marketplace.

The company has released the model under an MIT license on GitHub and Hugging Face, allowing commercial use without additional disclosure requirements. Pricing for the commercial API is $0.75 per million input tokens and $2.95 per million output tokens, with promotional discounts. Meituan’s move follows its 2023 acquisition of the AI startup Light Year Beyond and reflects a broader Chinese strategy to develop high‑performance AI capabilities without reliance on foreign silicon.