Skip to main content

Meituan Unveils LongCat-Flash-Thinking AI Model

Meituan Breaks New Ground with LongCat-Flash-Thinking AI Model

Chinese tech giant Meituan has unveiled LongCat-Flash-Thinking, a revolutionary reasoning model based on mixture-of-experts architecture. With 56 billion parameters, this model represents a significant leap forward in artificial intelligence capabilities.

Image

Dynamic Architecture Delivers Efficiency

The model's standout feature is its dynamic parameter activation, which adjusts between 18.6 billion and 31.3 billion parameters depending on context requirements, averaging about 27 billion active parameters. This innovative approach allows for:

  • Optimized computational efficiency
  • Context-aware performance tuning
  • Reduced energy consumption
  • Scalable task handling

Benchmark Dominance Across Multiple Disciplines

Meituan's official testing data reveals exceptional performance across critical benchmarks:

Mathematical Reasoning:

  • MATH500: Top-tier accuracy scores
  • AIME25: Industry-leading results with native tools reducing token consumption by 64.5%

Other Key Areas:

  • General reasoning (GPQA-Diamond)
  • Code generation (LiveCodeBench, OJBench)
  • Formal theorem proving (MiniF2F-Test)

The model demonstrates particular strength in logical operations and complex problem-solving scenarios.

Image

Open Access for Developers

In a move that will accelerate AI development:

  • Model weights released under MIT license
  • Detailed chat templates provided
  • Dedicated chat website launched (longcat.chat)

The open-source approach allows researchers worldwide to build upon Meituan's work while maintaining commercial flexibility.

Strategic Implications for AI Development

The launch positions Meituan as:

  1. A serious contender in foundational AI research
  2. An innovator in efficient large-scale models
  3. A contributor to the global AI ecosystem through open sourcing

The company emphasizes this release represents more than technological achievement—it's a strategic investment shaping future AI applications.

Key Points:

  • 🚀 56B parameter mixture-of-experts model launched
  • ⚡ Dynamic activation of 18.6B–31.3B parameters
  • 🏆 Outperforms competitors in math/logic benchmarks
  • 🔓 Fully open-sourced under MIT license
  • 🌐 Dedicated chat interface available