The food-delivery company just dropped a frontier-scale LLM. LongCat-2.0 is a Mixture-of-Experts model with 1.6 trillion total parameters, ~48B activated per token, and a 1M-token context window. The whole pretraining run — 35T+ tokens — happened on 50,000+ domestic AI ASICs in superpod clusters. NVIDIA wasn’t in the building.
What it actually is
This is an open-weight model built for agents, not chat. The deep optimizations target multi-step task planning, tool calling, and production-grade code generation — Meituan pitches it for “complex task planning and enterprise automation.” Under the hood: sparse attention, N-gram embeddings, zero-computation experts, and post-training specialist groups for coding and agent workflows.
Why it matters
The weights are downloadable on Hugging Face, so you run it yourself or wire it into an agent stack — no rate limits, no vendor lock. But the real headline is geopolitical: a Chinese company trained a trillion-parameter open model entirely on home-grown silicon, under export controls, with no major training rollbacks. That’s the proof-of-concept everyone was waiting for. The context doubled, the scale doubled, and the chip dependency went to zero.
You Might Also Like
- Moonshine Open Weights stt the Tiny Speech Model That Punches way Above its Weight
- Hermes Agent by Nous Research Might be the Open Source ai Agent That Finally Remembers Everything
- Nvidia Nemotron 3 Super 120b Parameters 12b Active the Math Behind the Fastest Open Source Reasoning Model
- Langchains gtm Agent Drove 250 More Conversions now the Framework Behind it is Open Source
- Tempo Machine Payments Protocol 500m 100 Partners and a Bold Play to own the ai Agent Economy
Top AI Product 
Leave a comment