Meituan's LongCat-2.0-Preview Opens Testing: Trillion-Parameter Model Trained on Domestic Computing Power

Meituan's new generation basic large model LongCat-2.0-Preview has opened for testing, with a total parameter scale exceeding one trillion, ranking among the world's top large models. On the same day, DeepSeek released its new generation V4 large model, whose total and active parameter counts are basically the same as LongCat-2.0-Preview.

A bigger breakthrough of Meituan's model is that its entire training and inference process relies on domestic computing clusters. The training phase used 50,000 to 60,000 computing cards, which is the largest-scale large model training task completed on domestic computing power so far.