Meituan has released LongCat-2.0, a 1.6 trillion parameter Mixture-of-Experts model designed for agentic coding tasks. The architecture dynamically activates between 33 billion and 56 billion parameters per token, ensuring computational efficiency at scale.
The model features a context window of 1 million tokens, enabling it to process entire large codebases simultaneously. It achieved a score of 59.5 on the SWE-bench Pro benchmark and 70.8 on Terminal-Bench, demonstrating strong performance on real-world software engineering challenges. The model weights are publicly available on Hugging Face.
Critically, the entire training process was executed on a 50,000-card cluster of Chinese-manufactured hardware, completely bypassing restricted Nvidia H100 or AMD MI300X chips. This represents the first compute cluster of its scale built with domestic Chinese technology.
LongCat-2.0 nearly triples the parameter count of its predecessor, LongCat-Flash (560B), which launched in September 2025. The rapid progression signals a significant shift in AI infrastructure independence, potentially eroding the strategic leverage of US semiconductor export controls as Chinese firms achieve frontier capabilities using domestic chips.