智谱(02513)GLM-5发布：技术全面升级 Agent能力达开源SOTA

Core Insights - The article highlights the launch of the new flagship model GLM-5 by Zhipu, which is designed to perform complex system engineering and long-range agent tasks, showcasing state-of-the-art (SOTA) capabilities in agentic engineering, comparable to Claude Opus 4.5 [1] Group 1: Model Capabilities - GLM-5 represents a shift in the AGI industry from "Vibe Coding" to "Agentic Engineering," evolving from simple dialogue and rapid prototyping to autonomously solving real-world long-range system engineering challenges [1] - The model features significant technical advancements, including an expanded parameter scale of 744 billion and pre-training data of 28.5 trillion [1] - A new asynchronous reinforcement learning infrastructure called "Slime" has been developed to maximize the model's potential, along with the first integration of a sparse attention mechanism that reduces deployment costs while maintaining long text performance [1] Group 2: Benchmark Performance - In benchmark tests, GLM-5 achieved programming capabilities aligned with Claude Opus 4.5, scoring 77.8 and 56.2 in SWE-bench-Verified and Terminal Bench 2.0 respectively, marking the highest scores among open-source models and outperforming Gemini 3 Pro [1] - GLM-5 also demonstrates open-source SOTA agent capabilities, achieving top performance in BrowseComp (networked retrieval and information understanding), MCP-Atlas (tool invocation and multi-step task execution), and τ²-Bench (planning and execution in complex multi-tool scenarios) [2]