China's Meituan has introduced LongCat-2.0, the country’s biggest AI language model fully trained and inferred using homegrown chips, demonstrating a significant breakthrough in China's domestic AI computing power capabilities.
- LongCat-2.0 is China's largest AI model trained fully on local hardware.
- Model has 1.6 trillion parameters and a 1 million token context window.
- Meituan overcame significant memory and infrastructure challenges.
What happened
China-based on-demand service Meituan released LongCat-2.0, a large language model with 1.6 trillion parameters and an exceptionally long 1 million token context window. The company open-sourced the model, which was fully trained and run on domestic AI chips across a 50,000-card computing cluster. This marks the first time a trillion-parameter LLM has completed both pre-training and inference exclusively on Chinese-made hardware.
LongCat-2.0’s training involved large-scale clusters of ASIC superpods designed specifically for AI workloads. Meituan utilized Huawei’s Collective Communication Library (HCCL) for enhanced training stability. This contrasts with previous domestic models that used local chips only for inference, making LongCat-2.0 a major step in proving the viability of China’s AI chip ecosystem for intensive pre-training tasks.
Why it matters
The launch demonstrates significant progress in China’s strategic push for technological self-reliance in AI, reducing reliance on Western GPUs such as those from Nvidia, which face export restrictions. LongCat-2.0’s scale rivals other advanced models like DeepSeek’s V4-pro, positioning Chinese AI development closer to global frontiers despite remaining behind leading models like OpenAI’s GPT-5.5.
Successfully conducting frontier-scale AI training on domestic chips challenges long-held assumptions about their limitations, particularly in memory capacity and software support. It underscores a maturing domestic hardware and software ecosystem capable of tackling industry-scale AI workloads, crucial for national tech sovereignty and innovation.
What to watch next
Further evaluation of LongCat-2.0 on leading AI benchmarks and advanced testing suites will provide clearer insight into its performance relative to global models. Meituan’s approach to overcoming hardware limitations through infrastructure optimization also suggests ongoing innovation needed to close the gap with established GPU ecosystems.
Industry observers will closely monitor how this development influences China’s AI chip market, including potential collaborations, competition, and adoption beyond tech giants. Huawei’s role and advancements in communication libraries like HCCL will also be key to scaling up future AI model training and deployment domestically.