Meituan Quietly Tests Trillion-Parameter AI Model Built on Domestic Compute

Meituan has started limited testing of a new large language model with more than 1 trillion parameters.

The model was reportedly trained on domestically developed compute clusters, likely without NVIDIA GPUs.

This follows Meituan’s earlier 560B-parameter LongCat-Flash release and broader AI investment push.

The test highlights China’s progress in training frontier models on local hardware and reducing reliance on NVIDIA.