Meituan, the Chinese food delivery and services giant, has released a new open-source large language...
The AMW Read
Novelty 3: overturns assumption that frontier models require Western chips, introducing a new class of domestically-trained model. Significance 2: structurally shifts compute supply narrative for China's AI ecosystem, with segment-level but not cross-segment adoption proof yet.
Meituan, the Chinese food delivery and services giant, has released a new open-source large language model trained exclusively on Chinese-made chips. This marks the first known instance of a major Chinese consumer internet company fully training a frontier-scale model without reliance on Western semiconductor hardware.
Why it matters: The model directly validates the progress of China's domestic AI chip ecosystem, a key geopolitical and supply-chain concern as US export controls tighten. It also signals that Chinese enterprises can achieve competitive model performance on alternative silicon, potentially reshaping the compute economics for the country's AI sector. The open-source release further aligns with China's strategy of distributing sovereign AI capabilities through widely accessible open-weight models.
Expert take: This event updates the player map in the foundation model segment — Meituan enters as a model developer, not just a consumer. More critically, it provides a real-world stress test for the viability of Chinese chips like those from Huawei (昇腾) or Cambricon (寒武纪) at scale. If this model's inference and training efficiency prove viable in production, it could accelerate a broader decoupling of China's AI supply chain from NVIDIA hardware, with deep implications for the global compute market and export control dynamics.
