Meituan Open-Sources 1.6T AI Model Trained 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 1.6T-parameter MoE model with ~48B active parameters and a 1M-token context window built around LongCat Sparse Attention.
- LongCat-2.0 was trained on a 50,000-chip cluster of domestic Chinese processors (described as AI ASIC superpods), with more than 30 trillion tokens of pretraining data, Meituan said.
- LongCat-2.0 reportedly outperformed GPT-5.5 on SWE-bench Pro and had been leading OpenRouter under the codename "Owl Alpha" before its public release, according to coverage by VentureBeat and commentators.
- Meituan — China's largest food delivery platform — built and deployed the model without Nvidia GPUs, a point underscored by multiple outlets (The Decoder, VentureBeat) and analysts citing Jensen Huang's export-controls argument.
- The release was framed across Reuters, SCMP and SiliconANGLE as China's largest open AI model trained on local chips, intensifying the narrative that US chip restrictions are redirecting — not capping — Chinese frontier-model development.
Why it matters: Chinese delivery giant Meituan just shipped a 1.6-trillion-parameter open model trained on 50,000 domestic chips, beating GPT-5.5 on a coding benchmark and leading OpenRouter — concrete proof that Nvidia-independent frontier training is no longer hypothetical, which directly pressures the premise of US GPU export controls.
