Meituan Unveils LongCat-2.0 AI Model Trained on Domestic Chinese Chips Amid U.S. Export Controls

Meituan has launched LongCat-2.0, a 1.6-trillion-parameter AI model trained entirely on domestic Chinese ASIC superpods, circumventing U.S. export restrictions on Nvidia's CUDA-based chips. The sparse mixture-of-experts architecture enables cost-efficient inference by activating only specialized sub-models per token, mirroring designs from DeepSeek and Mistral's Mixtral.

The model boasts a one-million-token context window—far exceeding benchmarks set by OpenAI's GPT-OSS and DeepSeek-R1-0528—though third-party validation of its performance against Google, OpenAI, and Anthropic's closed-source models remains pending. Meituan positions LongCat-2.0 as the backbone for AI agents and coding tools, targeting repository-wide edits and automated task execution.

Bernstein estimates Nvidia still commands 40% of China's AI chip market by 2025, but LongCat-2.0 signals a strategic pivot toward self-reliance. 'Frontier-scale AI without Western hardware isn't just aspirational—it's operational,' a Meituan spokesperson remarked.