SuperAPI AI Directory
Partner with us中文AI newsFavorites
DirectoryNews and tutorials

LongCat - Meituan's self-developed generative AI large model

Author: AI小集 · Original source: AI工具集

What is LongCat

LongCat(LongCat) is a generative AI large model independently developed by Meituan, designed to improve the company's internal work efficiency and innovation capabilities through artificial intelligence technology. The model has strong multimodal capabilities and can handle multiple data types such as text and images, performing excellently on generative tasks, for example quickly generating images, video scripts, code, and more. LongCat has been integrated into six major office scenarios within Meituan, including AI programming, smart meetings, and document assistants, effectively improving employee work efficiency.

Meituan has open-sourced the "LongCat-Flash" large model, with 560 billion parameters, adopting a MoE architecture and dynamic computation mechanism. LongCat-Flash can automatically adjust computational resources according to task complexity, with the number of activated parameters dynamically adjusted between 18.6 billion and 31.3 billion, averaging about 27 billion. A PID controller adjusts expert bias to ensure the average activated parameters remain stable at around 27 billion, maintaining stable computational load. In multiple benchmark tests, LongCat-Flash performs excellently. Compared with top models such as GPT-4.1, Claude4 Sonnet, and Gemini2.5 Flash, it is competitive in multiple fields, especially achieving SOTA level on agent tasks.

LongCat

Main features of LongCat

  • Intelligent programming assistance: LongCat can assist programmers with code generation and debugging, providing code snippet suggestions, logic optimization plans, and error troubleshooting guidance, helping developers quickly complete programming tasks and improve development efficiency.
  • Smart meeting assistant: In 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 organizing meeting records.
  • Document generation and editing: LongCat can quickly generate document content based on the topic or outline input by the user, provide intelligent writing suggestions, optimize document structure and language expression, and help users efficiently complete document writing and editing work.
  • Graphic design and video production: LongCat can generate images and video scripts, providing designers and video creators with creative inspiration and detailed storyboard design, assisting in the creation of high-quality visual content, and improving design and production efficiency.
  • Intelligent Q&A and knowledge management: LongCat can answer various user questions and provide accurate information and solutions. It can help organize and manage knowledge bases, making it convenient for users to quickly find and learn relevant knowledge.
  • Multimodal content generation: LongCat has strong multimodal capabilities and can handle multiple data types such as text and images, generating corresponding output content, such as generating images based on text descriptions, or generating text descriptions based on images, meeting diverse business needs.

Technical principles of LongCat

  • Transformer architecture: Achieves semantic encoding of input text and generation of output text through multi-layer encoders and decoders.
  • Large-scale pre-training: The model is pre-trained on massive text data to learn the general rules and patterns of language. The pre-training stage uses unsupervised learning, enabling the model to understand the grammar and semantics of natural language.
  • Hybrid training strategy: When training "LongCat", Meituan combined open-source data and internal private data, enabling the model to better adapt to Meituan's business scenarios.
  • Optimizing computational utilization: Drawing on the research approach of DeepMind's Chinchilla model, under a fixed compute budget, it improves the model's training efficiency by reasonably allocating model size and training data volume.
  • Fine-tuning and Application: On the basis of pretraining, fine-tuning is carried out for specific business scenarios and tasks to further improve the model's performance and adaptability.

LongCat Project Address

  • Official Website Address：https://longcat.chat
  • HuggingFace Model Address：https://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

LongCat Application Scenarios

  • AI Programming: LongCat can assist programmers with code generation and debugging, improving programming efficiency.
  • Smart Meetings: LongCat can automatically generate meeting minutes and extract key information, helping participants better understand and follow up on meeting content.
  • Document Assistant: LongCat can help employees quickly generate and edit documents, providing intelligent writing suggestions and content generation.
  • Graphic Design and Video Production: LongCat can generate images and video scripts, providing creative inspiration for designers and video creators.