Meituan LongCat is the AI division of the Chinese technology company Meituan, incorporated in the People's Republic of China and headquartered in Beijing. Meituan was founded in 2010 by Wang Xing as an e-commerce platform and food delivery service. The LongCat brand (known in the Chinese community as "Long Mao" or "Dragon Cat") was officially launched in 2025 to position the company in the large language and multimodal model market. LongCat champions an approach to building efficient AI models through architectural innovations and high-quality data, rather than simply scaling parameter counts.
In August 2025, LongCat introduced LongCat-Flash-Chat — a 560-billion parameter MoE model with dynamic expert activation ranging from 18.6 to 31.3 billion parameters depending on the task. This was followed in September by LongCat-Flash-Thinking — a specialized reasoning model version trained using a "cold start" method with a large-scale agent-based Reinforcement Learning approach. These models demonstrated superior performance on benchmarks, particularly in programming tasks.
In October 2025, LongCat released the LongCat-Video model for generating long (up to 5 minutes) and stable videos based on a Diffusion Transformer (DiT) architecture utilizing a Hierarchical Token Compression (HiCo) strategy. Then, in December 2025, LongCat-Image was launched — a 6-billion parameter image generation model that surpassed larger contemporaries at the time in photorealism and text rendering accuracy. It addressed a chronic issue in generative AI: poor rendering of non-Latin fonts.
Meituan LongCat actively engages with the global open-source community: all key models are released under the MIT license and are available on Hugging Face and GitHub. The company not only provides model weights and checkpoints but also publishes detailed technical reports, making their approaches reproducible.