Meituan Ships 1.6T Model on 50K Chinese Chips
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Meituan (China's largest food delivery platform, 3690.HK) open-sourced LongCat-2.0, a 1.6T-parameter MoE model with ~48B active parameters, trained entirely on a 50,000-chip cluster of Chinese domestic accelerators — no Nvidia GPUs involved.
- LongCat-2.0 uses a "LongCat Sparse Attention" (LSA) architecture to natively support a 1M token context window, and was built from the ground up for agentic coding tasks per the project's own launch post.
- Per AI commentators including @fellmentke and @meituan_longcat, LongCat-2.0 had already been leading OpenRouter anonymously under the codename "Owl Alpha" prior to release and reportedly outperforms GPT-5.5 on SWE-bench Pro.
- The model was trained on 30T+ tokens of pretraining data, per X commentator Poe Zhao — one of the largest publicly disclosed training runs on non-Nvidia silicon.
- Meituan's release reinforces X commentator Yuchen Jin's reading of Jensen Huang's Dwarkesh podcast warning that US export controls on Nvidia GPUs only accelerate the development of AI running on Chinese chips rather than containing it.
Why it matters: Meituan — China's largest food-delivery company, not a lab or big-tech incumbent — built a 1.6T-parameter model on 50,000 Chinese chips with zero Nvidia GPUs, validating Jensen Huang's own stated warning that US chip export controls accelerate rather than contain Chinese AI. The model reportedly matches frontier coding benchmarks, which means non-Nvidia silicon now competes at the top tier of agentic coding, not just in inference or fine-tuning.
