未知机构：国金计算机科技GLM5技术解析国产模型进入算力换效果阶段Token消耗-20260224

Summary of Key Points from the Conference Call Company and Industry Overview - The conference call discusses advancements in the domestic AI model, specifically focusing on the GLM-5 technology developed by Guojin Computer & Technology, which marks a significant evolution in the AI industry in China [1][2]. Core Insights and Arguments - Parameter Expansion: GLM-5 has doubled its total parameter count to 744 billion, with 40 billion active parameters, compared to the previous version GLM-4.5, which had 355 billion total parameters and 32 billion active parameters. This expansion represents a substantial increase in capacity [1]. - Performance Improvement: The model has shown an average improvement of approximately 20% across various core benchmark tests, positioning its overall capabilities on par with Claude Opus 4.5 and GPT-5.2. In specific tests, GLM-5 scored 77.8% in SWE-benchVerified and 75.9% in BrowseComp [1]. - Cost Efficiency: The GLM-5 model utilizes a DSA sparse attention architecture, which reduces GPU attention computation costs by half when processing long sequences. Additionally, it is optimized for domestic chip ecosystems, achieving performance comparable to international dual-GPU clusters while cutting deployment costs by 50% in long-sequence scenarios [2]. - Interleaved Thinking: The introduction of "Interleaved Thinking" allows for deep reasoning before each response and tool invocation, which is expected to lead to exponential improvements in computational efficiency [2]. - Shift to Agentic Engineering: GLM-5 aims to transition AI from passive code generation to autonomous planning and iterative "Agentic Engineering." Internal testing on the CC-Bench-V2 dataset has demonstrated strong end-to-end processing capabilities, indicating that the domestic model's capabilities have reached a level suitable for industrial applications [2]. Other Important Insights - Token Utilization: The model's ability to handle token consumption has significantly improved, suggesting a potential for increased scalability and application in various industrial contexts. The anticipated growth in token volume and international expansion is expected to benefit the model's adoption [2].