Meituan, China’s leading on-demand food delivery and local services giant, has unveiled LongCat-Flash-Chat, its own open-source large language model (LLM), marking a strategic move into the AI space.
The model, now available on Hugging Face, GitHub, and Meituan’s official site, boasts 560 billion parameters and uses a Mixture-of-Experts (MoE) architecture, which splits tasks among specialized sub-networks for faster, more efficient processing. Benchmark tests indicate LongCat-Flash-Chat performs on par with DeepSeek V3.1, Alibaba Cloud Qwen3, Moonshot AI Kimi K2, and even US counterparts like Anthropic’s Claude Sonnet and Google Gemini 2.5 Flash.
Notably, the model achieves over 100 tokens per second in inference speed at a cost as low as 5 yuan (~70 cents) per million tokens, highlighting efficiency alongside performance.
This release is Meituan’s first major salvo against Alibaba Cloud’s Qwen family, the world’s largest open-source AI ecosystem, and reflects the company’s broader strategy of aggressive AI adoption, building on prior AI products like NoCode, Kangaroo Consultant, and Meituan Jibai.
CEO Wang Xing emphasized a proactive approach: “AI will disrupt all industries, and our strategy is proactive offence rather than passive defence.”
The launch illustrates how Chinese AI firms are closing the gap with global peers, leveraging open-source models to drive innovation, collaboration, and market reach — even as Meituan’s core delivery business faces profit pressures.
