Meituan's LongCat-2.0: A Trillion-Parameter Open-Source AI Model Challenges Nvidia with Domestic Hardware

Meituan's LongCat-2.0, a 1.6 trillion-parameter Mixture-of-Experts (MoE) model with a native 1-million-token context window, has been officially released, immediately setting a new benchmark for open-source AI innovation. This formidable model boasts 1.6 trillion total parameters, yet intelligently activates only around 48 billion per token through its efficient MoE architecture. Crucially, LongCat-2.0 is the first trillion-parameter model known to complete its entire training and inference cycle on a vast 50,000-card cluster of domestic Chinese AI ASICs, explicitly sidestepping Nvidia hardware.

The model's native 1-million-token context, powered by the innovative LongCat Sparse Attention (LSA), dramatically improves efficiency by reducing computational complexity from quadratic to linear scaling for extended sequences. This immense context window is a significant leap for AI agents, enabling them to ingest entire codebases, voluminous documents, or complex conversational histories in a single pass, thereby maintaining full situational awareness for intricate, multi-step tasks. LongCat-2.0, purpose-built for agentic coding, has demonstrated strong performance, reportedly outperforming GPT-5.5 on the SWE-bench Pro benchmark. It had also quietly garnered significant traction as "Owl Alpha" on OpenRouter prior to its official reveal, topping usage charts.

Released under a permissive MIT license, LongCat-2.0 signals Meituan's bold pivot from a food delivery giant to a serious contender in foundational AI infrastructure. This development directly challenges the long-held assumption that frontier AI models are inseparable from Nvidia's ecosystem, showcasing the rapid advancements and strategic independence emerging from alternative hardware ecosystems. Its open nature and efficiency are poised to accelerate the development of more capable and cost-effective AI agents globally.