Chinese delivery app Meituan has unveiled an artificial intelligence model it claims outperforms ChatGPT and Claude. It is the largest model trained using only domestic hardware, without US-made chips. Amid the US-China technology rivalry, China appears to have taken a major step toward building its own AI supply chain.
According to the South China Morning Post (SCMP) on Monday, Meituan the previous day open-sourced its large language model (LLM) "LongCat-2.0." Meituan claimed the model matched or surpassed major commercial models including Google's Gemini, OpenAI's GPT-5.5, and Anthropic's Claude Opus in some coding and agent benchmarks. The model targets "agentic coding," and its preview version has already ranked among the three most-used models on the global AI marketplace OpenRouter, the company said.
LongCat-2.0 has 1.6 trillion parameters and a context window of 1 million tokens. This is comparable to "V4-pro," DeepSeek's latest flagship model released in April. Meituan claimed LongCat-2.0 is the industry's first trillion-parameter-class model to complete the entire training and inference process on a 50,000-unit domestic computing cluster.
The key is "pre-training." DeepSeek's V4-pro used domestic chips only for "inference." By contrast, LongCat-2.0 used domestic hardware for both inference and pre-training, according to Meituan. Pre-training is the process by which an AI model learns basic patterns from vast amounts of data, which carries a far greater computational burden.
Meituan said it built the model on a "large-scale cluster of tens of thousands of AI application-specific integrated circuit (ASIC) super pods." Unlike general-purpose processors, ASICs are chips made for specific tasks. The company did not specify its hardware supplier. However, in a WeChat post that day, the company said it used Huawei's Collective Communication Library (HCCL) to improve training stability. HCCL is a chip-to-chip communication system similar to Nvidia's collective communication library between GPUs (NCCL).
China's LongCat-2.0 Achieves First Pre-Training Success on Domestic Chips…"Memory Remains a Bottleneck"
Meituan, dubbed "China's version of DoorDash," is a latecomer in China's AI market. Its competitors include DeepSeek and ByteDance's Doubao. The LongCat team, established in 2023, released its first model only late last year.
Chinese domestic AI chips have been widely used for inference amid the country's technology self-reliance policy. However, domestic hardware has long been considered unsuitable for LLM pre-training. Most Chinese models trained on domestic chips have been small or limited to multimodal tasks. But the fact that LongCat-2.0 was trained solely on Chinese-made AI chips shows the growing importance of self-reliance in China's AI market. Major players such as DeepSeek, Alibaba, and ByteDance have been working to reduce their dependence on US-made chips since Washington's export controls in 2022.
The achievement drew immediate attention from industry insiders. Renowned technology analyst TP Huang said on X that day that it "dispelled concerns that Huawei's Atlas-950 super pods could not train large LLMs."
However, the company acknowledged the model still lags behind the world's top-tier models. Meituan said there remain high barriers to replacing Western GPUs. In its technical report, the company noted that "compared with the mature Nvidia GPU ecosystem, the supporting software community is still less developed." Pre-training on a cluster of more than 50,000 chips was "a significant system-level challenge due to the scale of the model and cluster," with memory being the "main bottleneck," the company explained. Meituan said the per-device memory of domestic accelerators falls significantly short of Nvidia's H800 chip. The H800 is a product banned from export to China under US regulations.
