Meituan Open-Sources 1.6T AI Model on Chinese Chips
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Meituan released and open-sourced LongCat-2.0, a 1.6 trillion-parameter MoE model with ~48B active parameters and a 1M-token context window, trained on a 50,000-chip cluster of domestic Chinese ASICs.
- LongCat-2.0 uses LongCat Sparse Attention (LSA) and was purpose-built for agentic coding — VentureBeat reports it is the full model behind the "Owl Alpha" release that had been leading OpenRouter.
- The Decoder headlined that the model "shows China can train massive AI models without Nvidia," while SCMP called it "the biggest AI model trained on local chips" — the cross-coverage consensus frames this as a milestone in Chinese compute self-sufficiency.
- An X post from Yuchen Jin (@yuchenj_uw) connected the release to Jensen Huang's stated view that US export controls on Nvidia GPUs won't stop China but will accelerate domestic-chip AI development.
- One technical observer (@fellmentke) noted LongCat-2.0 reportedly outperformed GPT-5.5 on SWE-bench Pro, though Meituan withheld full details on its training infrastructure.
Why it matters: A food delivery company producing a 1.6T-parameter model on 50K domestic Chinese chips reframes US export controls as an accelerant rather than a blockade — the exact point Jensen Huang was quoted as making. If non-AI-native Chinese firms can reach near-frontier performance on entirely domestic silicon, Nvidia's China revenue and the strategic premise of chip restrictions both erode simultaneously.
