Particle.news
Download on the App Store
4 ARTICLES
·
last wk.

Meituan Open-Sources LongCat-2.0, a 1.6 Trillion-Parameter LLM Trained on Chinese Chips

The release suggests China can run frontier-scale training on home-grown ASIC clusters and may reduce the leverage of US export controls.

A journalist checks the Starbucks menu on the Meituan app on her phone, during a media event of Starbucks launching a partnership with Meituan, at a Starbucks flagship store in Beijing, China January 18, 2022. REUTERS/Tingshu Wang/File Photo
Image
Image
Image

Overview

  • Meituan released and open-sourced LongCat-2.0 on Tuesday, June 30, 2026, publishing model weights and resources on Hugging Face.
  • The model has 1.6 trillion parameters, uses a Mixture-of-Experts design that activates about 33–56 billion parameters per token, and supports a 1 million-token context window for very long inputs.
  • Meituan says it trained and ran the model end-to-end on a 50,000-card domestic ASIC cluster and used Huawei’s HCCL for multi-chip communication and training stability.
  • Early benchmark results reported include a 59.5 on SWE-bench Pro and a 70.8 on Terminal-Bench, and Meituan said LongCat-2.0 matched or exceeded some leading proprietary coding models on certain tests.
  • The open release lowers barriers for developers to experiment with large coding models and could weaken US export-control pressure by showing domestic pre-training at frontier scale, which may shift competition and procurement choices in China’s AI sector.