TLDR: Chinese delivery giant Meituan has launched its open-source AI model, LongCat-Flash-Chat, featuring a Mixture-of-Experts (MoE) architecture with 560 billion parameters. The model, which activates 18.6 to 31.3 billion parameters at a time, excels in agent-based tasks and has been deployed internally to boost efficiency in various AI tools and customer service by 20%. This move positions Meituan in the competitive open-source AI landscape alongside major Chinese tech firms like Alibaba, Baidu, and ByteDance.
Chinese delivery giant Meituan has officially entered the burgeoning open-source artificial intelligence (AI) competition with the introduction of its new system, LongCat-Flash-Chat. The model was unveiled on Monday, September 1, 2025, and made available across platforms including GitHub, Hugging Face, and Meituan’s official website.
LongCat-Flash-Chat, affectionately nicknamed ‘LongCat’ (龙猫) in Chinese, is built upon a sophisticated Mixture-of-Experts (MoE) architecture. This design allows the model to boast a substantial 560 billion total parameters, yet it intelligently activates only a fraction—between 18.6 billion and 31.3 billion—at any given time. While Meituan acknowledges that LongCat-Flash-Chat does not currently possess advanced reasoning capabilities, the company asserts that it delivers performance comparable to leading AI models, particularly excelling in agent-based tasks, while utilizing fewer active parameters. This efficiency also translates into competitive output costs and faster generation speeds.
Meituan has been proactive in integrating its AI advancements internally. By early 2025, LongCat had already been deployed to power a suite of AI tools within the company, assisting with coding, meeting summaries, document preparation, and short-video production. Furthermore, the model has been instrumental in enhancing customer service and sales operations, leading to a reported 20% increase in support efficiency.
Also Read:
- China Achieves Landmark Generative AI Training Across Distributed Data Centers and Diverse GPU Architectures
- Alibaba’s Qwen Team Unveils Advanced Multi-Agent Framework for GUI Automation: GUI-Owl and Mobile-Agent-v3 Set New Benchmarks
The launch of LongCat-Flash-Chat places Meituan firmly within a crowded and rapidly evolving open-source AI landscape in China. Other prominent domestic tech players are also making significant strides in this domain. Baidu Inc., for instance, rolled out its Wenxin 4.5 model and the X1 inference model in March 2025. Alibaba Group Holding Ltd. launched its Qwen3 model in April 2025, with an update following in July. ByteDance Ltd. released its Doubao 1.5 reasoning model in April 2025, subsequently updating it to version 1.6 in June. Additionally, the Chinese startup Moonshot AI debuted its Kimi-K2 model in July 2025, which quickly ascended to the top of Hugging Face’s global open-source rankings, demonstrating strong performance in coding tasks.
