LongCat (Totoro) - Meituan's self-developed generative AI model

What is Totoro LongCat?

LongCat is a generative AI model independently developed by Meituan, designed to improve internal work efficiency and innovation through artificial intelligence technology. The model boasts powerful multimodal capabilities, handling various data types such as text and images, and excels in generative tasks, such as quickly generating images, video scripts, and code. LongCat has been integrated into six major Meituan office scenarios, including AI programming, intelligent meetings, and document assistants, effectively improving employee productivity.

Meituan has open-sourced its large-scale model, "LongCat-Flash," which boasts 560 billion parameters and employs a MoE architecture and dynamic computation mechanism. LongCat-Flash automatically adjusts computational resources based on task complexity, dynamically adjusting the number of activated parameters between 18.6 billion and 31.3 billion, averaging approximately 27 billion. By adjusting the expert bias through a PID controller, it ensures the average number of activated parameters remains stable at around 27 billion, maintaining a stable computational load. In multiple benchmark tests, LongCat-Flash performs exceptionally well, demonstrating competitiveness in several domains compared to top-tier models such as GPT-4.1, Claude4 Sonnet, and Gemini2.5 Flash, particularly achieving state-of-the-art (SOTA) performance in agent tasks.

LongCat's main functions

• Intelligent programming assistanceLongCat can assist programmers in code generation and debugging, providing code snippet suggestions, logic optimization solutions, and error troubleshooting guidance to help developers quickly complete programming tasks and improve development efficiency.
• Smart Meeting AssistantIn meeting scenarios, LongCat can automatically record meeting content, generate meeting minutes, extract key information and decision points, help participants better understand and follow up on meeting content, and save time in organizing meeting minutes.
• Document generation and editingLongCat can quickly generate document content based on the topic or outline entered by the user, provide intelligent writing suggestions, optimize document structure and language expression, and help users complete document writing and editing work efficiently.
• Graphic Design and Video ProductionLongCat can generate image and video scripts, providing designers and video creators with creative inspiration and detailed storyboard designs, assisting in the creation of high-quality visual content and improving design and production efficiency.
• Intelligent question answering and knowledge managementLongCat can answer various user questions, providing accurate information and solutions. It can also help organize and manage a knowledge base, making it easy for users to quickly find and learn relevant knowledge.
• Multimodal content generationLongCat boasts powerful multimodal capabilities, enabling it to process various data types such as text and images, and generate corresponding output content. For example, it can generate images based on text descriptions or generate text descriptions based on images, thus meeting diverse business needs.

The technical principle of LongCat Totoro

• Transformer architectureThe semantic encoding of the input text and the generation of the output text are achieved through a multi-layer encoder and decoder.
• Large-scale pre-trainingThe model is pre-trained on massive amounts of text data to learn the general rules and patterns of language. The pre-training stage uses unsupervised learning, enabling the model to understand the syntax and semantics of natural language.
• Hybrid training strategyWhen training the "Totoro" model, Meituan combined open-source data with internal private data to enable the model to better adapt to Meituan's business scenarios.
• Optimize computational utilizationThis approach draws inspiration from DeepMind's Chinchilla model, improving training efficiency by rationally allocating model size and training data within a fixed computational budget.
• Fine-tuning and applicationBased on pre-training, fine-tuning is performed for specific business scenarios and tasks to further improve the model's performance and adaptability.

LongCat's project address

• Official website addresshttps://longcat.chat
• HuggingFace model addresshttps://huggingface.co/meituan-longcat/LongCat-Flash-Chat
• Technical Report: https://github.com/meituan-longcat/LongCat-Flash-Chat/blob/main/tech_report.pdf
• Deployment Guide: https://github.com/meituan-longcat/LongCat-Flash-Chat/blob/main/docs/deployment_guide.md

Application scenarios of LongCat

• Artificial intelligence programmingLongCat can assist programmers in code generation and debugging, improving programming efficiency.
• Smart MeetingLongCat can automatically generate meeting minutes, extract key information, and help participants better understand and follow up on the meeting content.
• Document AssistantLongCat helps employees quickly generate and edit documents, providing intelligent writing suggestions and content generation.
• Graphic Design and Video ProductionLongCat can generate image and video scripts, providing creative inspiration for designers and video creators.