Meituan's 1.6T-Parameter LongCat-2.0 Tests China's Compute Independence Claim in Public

1. Meituan's 1.6T-Parameter LongCat-2.0 Tests China's Compute Independence Claim in Public

On June 30, 2026, Beijing-based Meituan released LongCat-2.0 as open-source weights. The model carries 1.6 trillion parameters and, according to Meituan, was trained entirely on domestic Chinese chips and compute clusters. Meituan is best known as a food delivery and local-services platform, which makes this release unusual in origin. The company has been quietly building out AI infrastructure, and LongCat-2.0 is the most visible product of that effort to date.

The domestic-chip claim is where the competitive weight sits. U.S. export controls on Nvidia H100s and A100s were designed to slow exactly this kind of large-scale model training in China. If Meituan's training run on domestic silicon is reproducible at this parameter count, it chips away at the assumption that export controls create a durable capability gap. That puts pressure on the policy case in Washington, and it puts Huawei's Ascend line and other domestic accelerators into a different conversation than they occupied six months ago. For Meta, which leads open-source weight releases at scale with Llama 4, LongCat-2.0 is a direct signal that the open-weight competitive landscape now has a second axis: not just capability, but hardware provenance.

The broader pattern is that Chinese technology conglomerates outside the core AI-lab category, Baidu, ByteDance, Alibaba, are now joined by a logistics-and-commerce platform in releasing frontier-scale open weights. Watch for independent evaluations of LongCat-2.0's actual benchmark performance, and watch for scrutiny of the domestic-chip training claim. Verification matters here. If the numbers hold, the export-control calculus shifts in ways that reach well beyond one model release.

Source: China's Meituan open-sources massive LongCat-2.0 AI model, saying it was trained on domestic chips