China's technology sector has taken a landmark step, demonstrating complete independence from American semiconductor technologies in the field of artificial intelligence. Meituan, a company known for its delivery services, has unveiled LongCat-2.0 — a large language model with 1.6 trillion parameters. The main sensation lies in the fact that the entire training cycle, from pre-training to inference, was carried out exclusively on domestic equipment, without using a single Nvidia chip.
This achievement places Meituan alongside advanced Chinese developments such as DeepSeek V4-pro. However, the key difference with LongCat-2.0 is that DeepSeek used local chips only for the final inference stage (responding to queries), whereas Meituan utilized domestic ASIC accelerators for the most resource-intensive stage — model pre-training.
Scale and Infrastructure
To train LongCat-2.0, a giant cluster of 50,000 domestic computing cards was deployed. The company did not disclose the direct supplier but confirmed the use of Huawei's communication library (HCCL) to ensure process stability. This dispelled doubts about the ability of Huawei Atlas-950 superclusters to handle training models of trillion-parameter scale. According to analyst TP Huang, this case removes all questions about the capabilities of Chinese "hardware" in this area.
Benchmarks and Competitive Position
LongCat-2.0 has already demonstrated impressive results. In tests related to programming and agent tasks (Terminal-Bench 2.1 and SWE-Bench Pro), the model outperformed an earlier version of Google Gemini 3.1 Pro. However, the company honestly admits that it still lags behind global flagships — OpenAI GPT-5.5 and Anthropic Claude 4.8 Opus. The model has not yet been evaluated on leading independent platforms such as Artificial Analysis and Arena, nor on the latest tests Agents’ Last Exam and CyberGym.
Challenges and Prospects
Despite the success, Meituan's engineers faced serious systemic difficulties. The main bottleneck was memory: domestic accelerators have significantly less memory per device compared to the Nvidia H800 chip, which is banned for export to China. To overcome this limitation and create a stable, scalable infrastructure, the company had to make tremendous efforts. The software ecosystem around Chinese "hardware" still noticeably lags behind the mature infrastructure of Nvidia CUDA.
Expert Opinion: The launch of LongCat-2.0 is not just a technological demonstration but a clear signal to the market. China has proven it can create competitive LLM models without relying on American chips. However, for investors and analysts, the key question remains the economic efficiency of such solutions. For now, the total cost of ownership and operation of clusters with 50,000 ASIC chips may be higher than that of comparable performance systems based on Nvidia. Nevertheless, this step significantly strengthens China's technological sovereignty and reduces risks associated with US export restrictions. This is a long-term trend that will reshape the global AI infrastructure map.