Meituan has open sourced its 560B-parameter LongCat-Flash-Chat model, aiming to boost global AI research with faster, cheaper, and more efficient agentic capabilities.
On September 1, Meituan announced the official release of LongCat-Flash-Chat, a large language model now open-sourced on GitHub, Hugging Face, and the company’s official website. The move marks Meituan’s first public contribution under its ‘Building LLM’ strategy, signalling its entry into the global open source AI ecosystem.
LongCat-Flash-Chat is built on a Mixture-of-Experts (MoE) architecture with 560 billion parameters, of which only 18.6 billion to 31.3 billion are activated per token, averaging 27 billion. This design achieves a dual optimisation of efficiency and performance, allowing the model to deliver on par with leading mainstream models while excelling in agentic tasks.
The model’s design introduces a ‘Zero-Computation Experts’ mechanism and employs a PID controller to stabilise activation levels. Inter-layer cross-channel pathways further parallelise MoE communication and computation, enabling training to be completed within 30 days. Inference efficiency is a highlight, reaching 100 tokens per second on H800 GPUs, with output costs as low as 5 yuan per million tokens.
Meituan also applied advanced techniques such as hyperparameter transfer, model stacking, and multiple stability strategies to ensure smooth training. For enhanced agentic capabilities, the company developed its own evaluation dataset and adopted multi-agent methods to generate diverse, high-quality trajectory data.
This release builds on Meituan’s broader AI strategy, which operates across three levels: AI at Work, AI in Products, and Building LLM. Previous AI applications include NoCode (AI coding agent), Kangaroo Advisor (business decision assistant), and Meituan Jibai (AI hotel management tool).
By open sourcing LongCat-Flash-Chat, Meituan highlights its focus on scalability, inference speed, and cost efficiency, making its innovations accessible to researchers and developers worldwide.
