The release of Meituan’s LongCat 2.0 (龙猫2.0) marks a significant milestone in China’s quest for technological self-reliance. As the first domestic large language model (LLM) with over a trillion parameters trained and deployed on a massive 50,000-card indigenous computing cluster, it challenges the long-held belief that cutting-edge AI remains impossible without Nvidia’s high-end silicon. For three years, Meituan quietly developed this infrastructure, proving that domestic hardware can now support the most demanding generative AI workloads.
While the model was tested anonymously under the pseudonym 'Owl Alpha' on the global API platform OpenRouter, its performance spoke for itself. By the end of June 2026, it had surged to become one of the top three most-called models globally, even securing the top spot in specific agent-based scenarios. This success suggests that the bottleneck for Chinese AI is shifting from hardware availability to software optimization and architectural efficiency.
Meituan’s CEO Wang Xing has reframed the company’s mission, moving beyond food delivery to position the firm as an 'offensive' AI powerhouse. Unlike competitors who treat AI as a defensive measure to protect existing market share, Meituan has integrated its AI Transformation (AIT) department into its core business structure. This strategic pivot reflects a broader sentiment among Chinese tech giants: in an era of prolonged geopolitical tension, 'sovereign compute' is no longer an option but a survival necessity.
The technical achievements of LongCat 2.0—including a 1.6 trillion parameter Mixture-of-Experts (MoE) architecture and a one-million-token context window—were made possible by overcoming the inherent instability of domestic chip clusters. By optimizing pipeline scheduling and memory usage, Meituan reportedly increased its model flops utilization (MFU) by 1.5 times. These marginal gains are critical in bridging the performance gap left by US export controls on advanced semiconductors.
Beyond Meituan, the broader Chinese ecosystem is showing similar resilience. From iFlytek’s specialized medical models to collaborative projects involving Huawei’s Ascend chips and Shenzhen’s research institutes, the trend toward a 'de-Nvidia-ized' supply chain is accelerating. The narrative in the Chinese AI sector is moving past the question of whether domestic chips are usable, focusing instead on how to make them the more cost-effective and stable choice for the next generation of intelligent applications.