Meituan's LongCat-2.0 Stretches to 1.6T Parameters Without Once Nudging Nvidia 🐱

Meituan has released LongCat-2.0, a 1.6 trillion-parameter open-source large language model trained entirely on domestic Chinese hardware, the Beijing-based food delivery and tech company confirmed in a June 30, 2026 announcement. The system is the first trillion-parameter model to complete both pre-training and inference on non-Nvidia silicon, Meituan said, distinguishing it from earlier domestic efforts such as DeepSeek's V4-pro, which relied on Chinese chips only for inference rather than the more compute-intensive training stage.

The model uses a Mixture-of-Experts architecture with roughly 48 billion active parameters, supports a 1-million-token context window, and introduces a custom mechanism called LongCat Sparse Attention designed to scale efficiently at that length. It is now available on OpenRouter under the name Owl Alpha. Meituan said the underlying cluster was built from large-scale ASIC superpods and orchestrated using Huawei's Collective Communication Library, or HCCL, to coordinate chip-to-chip traffic in a manner comparable to Nvidia's NCCL on GPU clusters.

On benchmarks, LongCat-2.0 outperformed Google's Gemini 3.1 Pro on Terminal-Bench 2.1 and SWE-Bench Pro, though Meituan acknowledged the model still trails global frontier systems including OpenAI's GPT-5.5 and Anthropic's Opus 4.8 on the most demanding agentic and reasoning evaluations. Industry analysts responded quickly, with tech analyst TP Huang writing that the launch should put to rest concerns about Huawei's Atlas-950 SuperPoDs, and Lehigh University researcher Hanchi Sun calling it the first model trained to near-frontier performance on 50,000 Chinese domestic accelerators.

External observers framed the release in geopolitical terms. Analyst Yuchen Jin wrote on X that the project echoes a remark by Nvidia chief executive Jensen Huang on the Dwarkesh podcast: "export controls on Nvidia GPUs won't stop China. They'll just accelerate the development of AI that runs on Chinese chips." Venture investors noted the milestone widens the global compute race, with one commenting that if frontier-scale training can run on local silicon at this level, the playing field is more open than at any previous point in the chip export-control era.