China's Meituan trains 1.6 trillion parameter LLM using 50,000 domestically sourced AI chips, demonstrating viable export-control bypass.

Meituan, the company best known for food delivery, released LongCat-2.0 on June 30—one of the most significant AI models developed in China. The model contains 1.6 trillion parameters and was trained entirely on a cluster of 50,000 domestically manufactured GPU chips, with no NVIDIA hardware involved.

The model uses a mixture of experts (MoE) architecture, where the 1.6 trillion total parameters include approximately 48 billion active parameters deployed per task. LongCat-2.0 was pretrained on over 30 trillion tokens and supports a context window of 1 million tokens. On the Terminal Bench 2 coding benchmark, it scored 70.8, placing it in competitive territory with leading global models. The 50,000-chip domestic cluster represents what appears to be the largest reported training run using non-NVIDIA hardware.

This development arrives as the US has progressively tightened export controls on advanced AI chips to China, operating under the assumption that restricting access to top-tier NVIDIA silicon would constrain Beijing's AI capabilities.

Separately, Moonshot AI has been advancing its Kimi model series, with variants reaching 1 trillion parameters trained on 15.5 trillion tokens. The Kimi K2 series features approximately 32 billion active parameters and incorporates a proprietary MuonClip optimizer designed to maintain stability during training. The research notes there is no direct association between Moonshot AI's efforts and the large domestic training cluster used by Meituan.