Recently, Meituan officially released its latest open-source large language model LongCat (Guinea Pig), aiming to promote the development and application of artificial intelligence technology. The release of this model not only marks another important advancement for Meituan in the AI field, but also provides developers and researchers with a powerful new tool.

The core version of LongCat LongCat-Flash has 56 billion parameters. Its core advantage lies in the innovative mix of experts (MoE) architecture. This architecture dynamically activates 18.6B to 31.3B parameters (approximately 27B on average) based on contextual needs, significantly optimizing computational efficiency. This enables LongCat to process over 100 tokens per second during inference, demonstrating excellent performance.

QQ20250901-091618.png

In terms of technical implementation, Meituan pays special attention to the training stability and scalability of the model. Through a series of optimization strategies, LongCat can efficiently train on a scale of over 100,000 accelerators and achieve low-latency inference. In addition, the model design considers dynamic allocation of parameter activation, ensuring consistent computational load across different tasks.

To enhance model capabilities, LongCat adopts a multi-stage training pipeline. The initial training aims to build a base model and optimize it through data fusion strategies; the later stage further enhances its reasoning and coding capabilities and expands the context length to 128k to meet more complex task requirements.

Evaluation results show that LongCat performs comparably to leading models in the industry across multiple benchmark tests, particularly demonstrating good accuracy in areas such as MMLU, mathematical reasoning, and general reasoning. These achievements not only prove the technical strength of LongCat, but also lay the foundation for its potential in practical applications.

Meituan stated that by opening up this powerful model, it hopes to promote the sharing and development of AI technology, provide convenience for more developers, and thus drive the progress of the entire industry. The open-source of LongCat will undoubtedly become a powerful assistant for future AI applications.

Address: https://longcat.chat/