2026-07-05 / Signal #4

LongCat-2.0: China’s 1.6T Open-Source MoE Agentic Coding Model Trained on 50K Domestic Chips (No Nvidia)

“The biggest AI model in the world is called LongCat, has 1.6 trillion parameters, and was built entirely without Nvidia chips.”

Why It Matters

Geopolitical bypass of Nvidia/US chip dominance via weirdly-named massive open model optimized for agents; changes builder tools, institutions, and compute economics with strange architecture tricks.

Evidence

Meituan LongCat official blog/longcatai.org (released June 30, 2026, trending HN), 1.6T params (~48B active), native 1M context, LongCat Sparse Attention, n-gram embeddings, trained on 30T+ tokens on domestic cluster for agentic coding/workflows.

Sources

Daily scan: 2026-07-05

No public source URL captured yet.