3 6 Ke·2025-09-02 08:15Core Viewpoint - Meituan is leveraging its "dispatch logic" from delivery services to the AI sector with the introduction of the LongCat-Flash model, aiming to optimize computational power usage and reduce costs in AI inference tasks [2][20]. Group 1: Technological Innovation - LongCat-Flash features a total parameter scale of 560 billion, but only a portion (approximately 18.6B–31.3B) is activated during inference, allowing for efficient resource allocation based on task complexity [2][5]. - The model incorporates "zero computation experts" to handle simple tasks directly, minimizing unnecessary computational expenditure and reserving resources for more complex tasks [3][5]. - The architecture includes a Shortcut-connected MoE (ScMoE) that allows for simultaneous task dispatch and processing, enhancing overall efficiency [6][8]. Group 2: Engineering Capability - LongCat-Flash's training approach resembles the gradual expansion of a delivery network, ensuring stability and efficiency before scaling up operations [9]. - The model employs a "threefold guarantee" system to prevent overload and ensure stable performance during operation [9]. Group 3: Performance Comparison - LongCat-Flash demonstrates competitive performance in various benchmark tests, achieving scores comparable to leading models in general tasks, complex reasoning, mathematical abilities, and programming tasks [10][14][16]. - In the MMLU benchmark, LongCat-Flash scored 89.71, while in CEval, it achieved 90.44, indicating strong capabilities in Chinese language understanding [10][14]. - The model's speed is highlighted as a significant advantage, with faster response times compared to competitors like Kimi 1.5 [16][18]. Group 4: Market Implications - Despite not having a clear edge in performance metrics, LongCat-Flash's speed and cost efficiency may disrupt the AI model market, as Meituan applies its operational strategies to AI [20]. - The company's approach of translating complex technological challenges into manageable logistics may provide a unique competitive advantage in the evolving AI landscape [20].