🌍Meituan's LongCat-2.0 Trains 1.6T Params on Chinese Chips
TL;DR
Meituan open-sourced LongCat-2.0 on June 30, a 1.6-trillion-parameter agentic coding model it says was both pre-trained and run entirely on domestic chips. The cluster used 50,000 home-grown accelerators and a 1M-token context window.
Meituan open-sourced LongCat-2.0 on June 30, a 1.6-trillion-parameter agentic coding model it says was both pre-trained and run entirely on domestic chips. The cluster used 50,000 home-grown accelerators and a 1M-token context window. It claims a first for full-process training off US silicon.
Key Points
1.6T parameters, 1M-token context, open-sourced June 30
Trained and inferred on a 50,000-card domestic cluster, Huawei HCCL in the stack
Claims first trillion-param model fully pre-trained on Chinese chips, unlike DeepSeek-V4-pro (inference only)
Positioned near-frontier and has been topping OpenRouter usage
Scale is on par with DeepSeek's V4-pro flagship
Why It Matters
If China can pre-train trillion-parameter models without Nvidia, US export controls lose their grip on the most strategic layer of the AI stack.
Quick Facts
Frequently Asked Questions
Why does this matter?
If China can pre-train trillion-parameter models without Nvidia, US export controls lose their grip on the most strategic layer of the AI stack.
What happened?
Meituan open-sourced LongCat-2.0 on June 30, a 1.6-trillion-parameter agentic coding model it says was both pre-trained and run entirely on domestic chips. The cluster used 50,000 home-grown accelerators and a 1M-token context window.
Comments
Be the first to comment
Enjoyed this article?
Get it daily. 7am. Free. Reads in 5 minutes.
Join 2,187 builders reading daily.