Meituan Trains 1.6T-Parameter AI Model Without Nvidia GPUs
Meituan releases LongCat-2.0, a 1.6-trillion-parameter foundation model built on a Mixture-of-Experts architecture with approximately 48 billion activated parameters per token. Before its official debut, the model reportedly appears on OpenRouter under the anonymous name Owl Alpha, where it climbs into the top three by total usage and ranks second globally in Claude Code Agent scenarios, trailing only Claude Opus 4.8. While the technical community views its agentic capability as roughly on par with Claude Opus 4.6, the parameter count alone does not make LongCat-2.0 the world's strongest model.
The real headline is what powers the model. According to Chinese analysis and language in Meituan's own model card, the entire training and deployment pipeline runs on domestic AI ASIC superpods with no Nvidia GPUs involved. The pretraining phase spans more than 35 trillion tokens and completes without rollbacks or irrecoverable loss spikes. This matters because China's previous domestic-compute achievements typically cover narrower milestones such as inference or post-training on local chips, whereas LongCat-2.0 represents a full trillion-parameter training-and-serving pipeline built entirely outside the Nvidia ecosystem.
For China's AI industry, this launch marks a significant geopolitical shift. If a major Chinese company can independently train and serve a model at this scale using only domestic hardware, it suggests that U.S. export controls on advanced GPUs may be less effective than intended. LongCat-2.0 may not top every benchmark, but it demonstrates that the gap between Nvidia-dependent AI development and fully homegrown alternatives continues to narrow at an pace that few analysts predicted.