LongCat

What is LongCat

LongCat is a self-developed large model launched by MeituanAI chat platform, with powerful natural language processing capabilities. The latest version, LongCat-Flash-Chat, is now officially open source, adopting an innovative Mixture-of-Experts (MoE) architecture, with a total parameter count of 560 billion, activating only 18.6 billion to 31.3 billion parameters per token, with an average of about 27 billion activated parameters, achieving efficient utilization of computing power. The LongCat-Flash-Chat model performs excellently in multiple benchmarks, especially showing clear advantages in agent tasks. For example, it surpasses other models in τ²-Bench, and ranks first in VitaBench with a score of 24.30. LongCat-Flash-Chat also performs excellently in programming, instruction following, and other areas, such as scoring 39.51 in TerminalBench, ranking second; and ranking first in IFEval with a score of 89.65.

LongCat's latest launchnativeThe "Deep Research" agent is an AI research assistant for local life scenarios, adopting a "search-analyze-render" multi-agent collaboration architecture, supporting 400 rounds of interaction and a 256K ultra-long context, and can automatically generate professional guide reports including price comparisons, route planning, and visual charts. The "Deep Research" agent is trained on Meituan's real POI data and the Rubrics evaluation system, emphasizes information accuracy and traceability, and can provide users with one-stop life decision-making services such as restaurant recommendations, travel planning, and hotel reservations.

LongCat's main features

• Natural language dialogue: Can conduct fluent and natural dialogue exchanges, accurately understand user intent, and provide precise and detailed answers, covering questions in many fields such as life, study, and work.
• Web search and real-time information integration: Supports web search functionality, can obtain the latest information on the internet in real time, integrate it into answers, and provide users with the latest data and information.
• Multi-domain knowledge application: Possesses knowledge in multiple fields such as science, technology, culture, history, and art, can provide users with broad information support, and answer various professional questions.
• Text generation and creation: Supports generation tasks for multiple text types such as article writing, story creation, and copywriting, and generates high-quality, logically clear text content according to user needs.
• Code generation and analysis: Can generate code snippets in multiple programming languages, analyze code, help users understand code logic and functionality, and is suitable for programming learning and development assistance.
• Agent task optimization: Performs excellently in agent tasks, possesses powerful tool use and complex scenario handling capabilities, and is suitable for application scenarios requiring intelligent decision-making and automated operations.
• Deep Research agent: Supports one-click generation of professional local life guide reports including search, analysis, and visualization, an intelligent decision-making assistant covering food, drink, entertainment, lodging, and travel.

How to use LongCat

• Visit the official website to experience it: You can directly visit the LongCat official website https://longcat.chat/ , start a conversation with LongCat-Flash-Chat, and perform operations such as natural language communication and obtaining information answers.
• Use the open-source model: Go to Hugging Face or Github to obtain the open-source model, and perform local deployment, secondary development, or research use according to your own needs.
• Standalone deployment experience: Using SGLang for single-machine deployment, run a specific command (such as the example command) to start the service, and you can use the LongCat-Flash-Chat model in your local environment.

LongCat's official website address

• Official website address：https://longcat.chat/
• Hugging Face model repository：https://huggingface.co/meituan-longcat/LongCat-Flash-Chat
• Github repository：https://github.com/meituan-longcat/LongCat-Flash-Chat

How is the LongCat model's performance

• Supported models: The LongCat dialogue platform currently mainly supports the LongCat-Flash-Chat model.
• Model evaluation performance：
  • In terms of general domain knowledge, LongCat-Flash achieved an excellent score of 86.50 on the ArenaHard-V2 benchmark, ranking second; it scored 89.71 on MMLU and 90.44 on CEval, demonstrating strong and comprehensive performance.
  • In terms of Agentic tool use, LongCat-Flash outperformed other models in τ²-Bench; it ranked first in VitaBench with a score of 24.30, demonstrating strong processing capability in complex scenarios.
  • In terms of programming, LongCat-Flash scored 39.51 in TerminalBench, ranking second; it scored 60.4 in SWE-Bench-Verified, demonstrating solid programming ability.
  • In terms of instruction following, LongCat-Flash ranked first in IFEval with a score of 89.65; it also achieved the best results in COLLIE and Meeseeks-zh, with 57.10 and 43.03 respectively, highlighting its excellent command of Chinese and English instruction sets.

• Model technical highlights:
  • Innovative Mixture-of-Experts model architecture: It adopts a Mixture-of-Experts (MoE) architecture, with a total parameter count of 560 billion, but each token activates only 18.6 billion to 31.3 billion parameters based on contextual needs, with an average activation of about 27 billion parameters, achieving on-demand allocation and efficient utilization of computing power.
  • Zero-Computation Experts mechanism: It introduces the "Zero-Computation Experts" mechanism, using a PID controller to fine-tune expert biases in real time, stabilizing the average activation per token at about 27 billion, effectively controlling total computing power consumption.
  • Cross-layer channel optimization: Cross-layer channels are laid between layers, enabling MoE communication and computation to be largely parallelized, greatly improving training and inference efficiency.
  • Efficient training and inference: Through customized low-level optimization, LongCat-Flash completed efficient training within 30 days and achieved an inference speed of 100+ tokens/s per user on H800.
  • Agent capability optimization: It built its own Agentic evaluation set to guide data strategy and optimized throughout the training process, including using multi-agent methods to generate diverse high-quality trajectory data, etc., significantly improving agent capabilities.

LongCat's application scenarios

• Intelligent dialogue and customer service: It provides smooth and natural conversational interaction and can serve as an intelligent customer service system, quickly and accurately answering user questions, improving user experience and customer service efficiency.
• Content creation and generation: It supports the generation of multiple text types such as articles, stories, and copywriting, helping creators quickly produce high-quality content and improving creation efficiency.
• Programming assistance: It generates code snippets in multiple programming languages and parses code logic, providing developers with programming suggestions and solutions, assisting programming learning and development work.
• Agent tasks: Performs excellently in complex scenarios, can be used for agent tasks such as automated operations and intelligent decision-making, and is suitable for scenarios that require efficient handling of complex tasks.
• Knowledge Q&A: Covers knowledge across multiple domains and can serve as a knowledge Q&A system, providing users with professional answers in fields such as science, technology, and culture.
• Educational Tutoring: Assists in the education field by providing learning materials and answering academic questions, helping students and educators learn and teach more efficiently.