Chinese food-delivery and services giant Meituan has open-sourced LongCat-2.0, a 1.6-trillion-parameter Mixture-of-Experts model built for agentic coding. Released on June 30 and published under an MIT license on Hugging Face, it is the first frontier-scale Chinese model to complete both training and inference end-to-end on a 50,000-chip cluster of domestic accelerators - without Nvidia hardware in the loop.
A trillion-parameter model that fits on domestic silicon
LongCat-2.0 activates only about 48 billion parameters per token on average - dynamically 33B to 56B depending on the input - through Meituan's "Zero-Computation Experts" and ScMoE routing. It supports a native 1 million token context via Long-Cat Sparse Attention (LSA), a linear-complexity sparse attention scheme.
Benchmarks: agentic coding leadership
The team reports 59.5 on SWE-bench Pro and 70.8 on Terminal-Bench, which it says leads Google Gemini 3.1 Pro, OpenAI's GPT-5.5 and Anthropic's Claude Opus 4.6 on deep software-engineering tasks. LongCat-2.0 is available on OpenRouter and has already climbed into the platform's global top three by call volume.
The chip story is the real story
Meituan trained the model on a 50,000-card cluster of domestic ASICs - a milestone in China's push to build frontier AI capacity without US chips. The move follows recent Chinese open-source releases from X Square Robot and comes as Washington tightens export controls on Western frontier models and considers offering a 5% equity stake to the US government.
Positioning against the West
Meituan is also undercutting Western frontier models on pricing, according to reporting from Crypto Briefing, positioning LongCat-2.0 as a cost-competitive agentic-coding alternative to OpenAI, Anthropic and Google offerings.
Reporting based on coverage from VentureBeat, LongCat AI, Crypto Briefing and KuCoin News.