China's First Self-Developed Trillion-Parameter Large Model for Card Training and Inference Officially Open Source - 36 Kr

Meituan has open-sourced its trillion-parameter large model, LongCat-2.0, according to a report by 36Kr. The model is described as China's first self-developed trillion-parameter large model for card training and inference to be officially open-sourced. LongCat-2.0 has 1.6 trillion total parameters and averages 48 billion activated parameters per inference. The model completed full-process training and inference on a 50,000-card domestic computing cluster.

The open-source release includes the model's weights and inference code. Meituan stated that the model achieves state-of-the-art performance on several benchmarks, though specific benchmark results were not detailed in the excerpt. The model is designed for both training and inference on domestic hardware, emphasizing self-sufficiency in AI infrastructure.

The 50,000-card cluster used for training is based on domestic chips, reducing reliance on foreign suppliers. This development may affect the global AI landscape by providing a high-capability open-source alternative to models from U.S. companies.