Loading...
LongCat
Meituan’s self-developed 560-billion-parameter MoE large-model dialogue platform that supports multimodal generation—including text, code, and images—providing developers and enterprises with fast and cost-efficient AI capabilities.
LongCat large modelMeituan LongCat-Flash-ChatMoE architecture open-source modelAI code generation toolagent task model
Features of LongCat
- 560-billion-parameter MoE architecture with an average of 2.7 billion active parameters and inference speeds up to 100 tokens/s
- Supports text generation, code completion, image creation, and multi-turn dialogue in a unified workflow
- Offers a web interface, APIs, and open-source weights for local deployment and secondary development
- Ranks first in evaluations such as VitaBench and τ²-Bench, delivering SOTA performance on complex agent tasks
Use Cases of LongCat
- Developers can quickly generate code snippets, debug scripts, and receive optimization suggestions
- Content teams can produce marketing copy, video scripts, and images at scale to shorten creative cycles
- Enterprises can build intelligent customer service and knowledge Q&A systems to reduce human agent costs
- Researchers can fine-tune the open-source weights to explore MoE training and inference optimization
FAQ about LongCat
Is LongCat AI free?
Online usage on the official website is free; API usage is billed at around 5 RMB per million tokens, and open-source weights can be deployed locally at no cost.
How do I deploy LongCat-Flash-Chat locally?
Clone the GitHub repository, load the Hugging Face weights using SGLang, and it can run on a single H800 GPU.
Which languages does the model support?
Primarily Chinese and English, with support for multilingual code comments and documentation generation; additional languages are continuously being expanded.
Is LongCat suitable for building multimodal agents?
Yes. It is optimized for tool usage and complex scenario decision-making, ranking first in the VitaBench evaluation.