About

Meituan Unveils LongCat-2.0, First Trillion-Parameter AI Model Trained Entirely on Domestic Chips

Published
Score
11

Why it matters

Meituan, the Chinese food delivery and services platform, launched LongCat-2.0 on June 30, a 1.6-trillion-parameter language model trained and deployed entirely on domestic Chinese semiconductors. The model features a one-million-token context window and claims performance parity with Google's Gemini 3.1 Pro. Critically, LongCat-2.0 was both trained from scratch and served on a 50,000-chip domestic compute cluster—distinguishing it from prior Chinese models like DeepSeek's V4-pro, which used domestic chips only for inference. Meituan has open-sourced the model weights.

The development represents a direct response to U.S. export controls restricting advanced semiconductor access to Chinese firms. Meituan's AI team began integrating domestic chips in 2023, suggesting a three-year path to frontier-scale training capability without NVIDIA hardware. The specific Chinese chipmakers supplying the AI ASICs remain unidentified in available disclosures.

For practitioners tracking AI regulation and export control enforcement, this milestone signals material progress in China's AI self-sufficiency strategy. The announcement aligns with parallel efforts by Baidu and Zhipu to migrate workloads to domestic silicon, indicating a coordinated industry shift away from U.S. chips. Attorneys monitoring sanctions compliance, foreign investment screening, or semiconductor export policy should treat this as evidence of accelerating Chinese capability in restricted technology domains—relevant context for CFIUS filings, ECRA enforcement actions, and broader geopolitical risk assessments in AI-adjacent sectors.

Sources

mail Subscribe to Artificial Intelligence email updates

Primary sources. No fluff. Straight to your inbox.

Also on LawSnap