China claims biggest AI model trained on local chips

LongCat-2.0: 1.6T parameters, 1M context. Meituan claims first trillion-parameter model with full-process training and inference on 50,000-card domestic cluster. Unlike DeepSeek-V4-pro (domestic chips for inference only), LongCat used domestic hardware for pre-training and inference. Built on ‘AI ASIC superpods’.