Skip to main content

Meituan Open-Sources LongCat AI Model to Boost Developer Innovation

Meituan Open-Sources Cutting-Edge LongCat AI Model

Chinese tech giant Meituan has officially released LongCat, its new open-source large language model, marking a significant stride in democratizing AI technology. The model's launch provides developers worldwide with access to advanced machine learning capabilities previously restricted to major corporations.

Technical Breakthroughs

The flagship LongCat-Flash version employs a 56 billion-parameter mixture of experts (MoE) architecture that dynamically activates between 18.6B to 31.3B parameters based on contextual requirements. This innovative approach delivers:

  • 100+ tokens processed per second during inference
  • Optimized computational efficiency through selective parameter activation
  • Consistent performance across diverse tasks

Image

Scalable Infrastructure

Meituan engineers prioritized training stability and scalability, enabling LongCat to:

  1. Train efficiently across 100,000+ accelerators
  2. Maintain low-latency inference speeds
  3. Automatically balance computational loads between tasks

The development team implemented a multi-stage training pipeline that progressively enhances the model's capabilities:

  • Phase 1: Base model construction with optimized data fusion
  • Phase 2: Enhanced reasoning and coding capabilities
  • Phase 3: Context window expansion to 128k tokens

Benchmark Performance

Independent evaluations demonstrate LongCat's competitive edge against industry leaders in:

  • MMLU (Massive Multitask Language Understanding)
  • Mathematical reasoning challenges
  • General problem-solving tasks

The model particularly shines in accuracy-intensive applications, making it suitable for enterprise-grade implementations.

Open-Source Strategy

By open-sourcing LongCat, Meituan aims to:

  • Accelerate AI innovation through community collaboration
  • Lower barriers to advanced ML development
  • Establish industry standards for efficient large models

The model is available at https://longcat.chat/

Key Points:

  • 🚀 56B parameter MoE architecture with dynamic activation
  • ⚡ Processes 100+ tokens/second during inference
  • 🧠 Supports extended 128k context windows
  • 📊 Matches leading models in benchmark tests
  • 🌍 Fully open-sourced for community development