Meituan Trains 1.6T-Parameter AI Model on Chinese Chips Alone

Chinese Chips Power Full AI Training Cycle

Food delivery giant Meituan has released what it describes as China's first trillion-parameter AI model trained completely on domestically produced chips, marking a significant step in the country's efforts to build advanced AI systems independent of foreign hardware.

The Beijing-based company open-sourced LongCat-2.0 on Tuesday, a large language model with 1.6 trillion parameters and a context window capable of processing 1 million tokens. The model's scale matches DeepSeek's V4-pro flagship, which launched in April.

Beyond Inference to Full Training

The key distinction lies in how the hardware was deployed. While DeepSeek's V4-pro used Chinese chips only for inference—the process of running a pre-trained model to answer queries—LongCat-2.0 relied on domestic hardware for both inference and the far more demanding pre-training phase.

Pre-training requires models to process massive datasets to learn fundamental patterns, consuming substantially more computational resources than inference. Successfully completing this phase on home-grown chips demonstrates a new level of capability for Chinese semiconductor technology.

Meituan built LongCat-2.0 on what it calls "large-scale clusters of tens of thousands of AI ASIC superpods." ASICs, or application-specific integrated circuits, are chips designed for particular workloads rather than general-purpose computing.

Huawei Connection Emerges

Though Meituan did not explicitly identify its chip supplier, the company revealed in a WeChat post that it used the Huawei Collective Communication Library (HCCL) to enhance training stability. HCCL is a chip-to-chip communication system analogous to Nvidia's widely used NCCL framework, strongly suggesting Huawei hardware underpinned the training infrastructure.

Why it matters

This release demonstrates that Chinese companies can now complete the entire AI development pipeline—from initial training through deployment—using only domestic chips. That capability reduces dependence on foreign technology at a time when export controls limit access to advanced processors from Nvidia and other Western manufacturers. For enterprises evaluating China's AI ecosystem, LongCat-2.0 shows that hardware constraints may pose less of a barrier to frontier-scale model development than previously assumed.

The details were first reported by the South China Morning Post.