Core Insights - The article highlights the launch of the new flagship model GLM-5 by Zhiyuan (02513), which is designed to handle complex system engineering and long-range agent tasks, showcasing state-of-the-art (SOTA) capabilities in agentic engineering, comparable to Claude Opus4.5 [1] Group 1: Model Capabilities - GLM-5 represents a shift in the AGI industry from "Vibe Coding" to "Agentic Engineering," evolving model capabilities from simple dialogue and rapid prototyping to autonomously solving real-world long-range system engineering challenges [1] - The model features a parameter scale expanded to 744 billion and pre-training data increased to 28.5 trillion [1] Group 2: Technical Innovations - GLM-5 incorporates a new asynchronous reinforcement learning infrastructure called "Slime," aimed at maximizing the model's potential [1] - The model integrates a sparse attention mechanism for improved long-text performance while significantly reducing deployment costs [1] Group 3: Benchmark Performance - In benchmark tests, GLM-5 achieved programming capabilities aligned with Claude Opus4.5, scoring 77.8 in SWE-bench-Verified and 56.2 in Terminal Bench2.0, marking the highest scores among open-source models and outperforming Gemini3Pro [1] Group 4: Agent Capabilities - GLM-5 also demonstrates open-source SOTA agent capabilities, achieving the top performance in BrowseComp (networked retrieval and information understanding), MCP-Atlas (tool invocation and multi-step task execution), and τ-Bench (planning and execution in complex multi-tool scenarios) [2]
智谱GLM-5发布：技术全面升级 Agent能力达开源SOTA