This is China's biggest AI model trained on local chips
What's the story
Chinese food delivery giant Meituan has unveiled LongCat-2.0, the country's largest artificial intelligence (AI) model trained entirely on home-grown hardware. The new large language model (LLM) comes with an impressive 1.6 trillion parameters and a context window of one million tokens. This puts it at par with DeepSeek's flagship model, V4-pro, which was launched in April this year.
Training details
LongCat-2.0 was pre-trained using domestically sourced hardware
Unlike DeepSeek V4-pro, which only used domestic chips for inference, LongCat-2.0 employed these home-grown hardware components for both inference and pre-training. Pre-training is a more computationally intensive process where an AI model learns basic patterns from massive datasets. Meituan claims that LongCat-2.0 was built entirely on "large-scale clusters of tens of thousands of AI ASIC superpods," demonstrating its capability to "conduct frontier-scale training on alternative hardware platforms."
Process
Meituan used Huawei's CCL for training
While Meituan didn't name its hardware supplier, the company did say that it used the Huawei Collective Communication Library (HCCL) to improve training stability. HCCL is a chip-to-chip communication system similar to NVIDIA's Collective Communication Library. This further highlights how advanced and sophisticated LongCat-2.0's training process was, using cutting-edge technology from one of China's leading tech companies.