智谱GLM-5开源发布，宣布“大幅降低”部署成本

Core Insights - The release of the new base model GLM-5 by Zhiyu Open Source marks a transition from the "Vibe Coding" era of writing code snippets to the "Intelligent Agent Engineering" era focused on completing system engineering [2][3] Group 1: Breakthrough Innovations - GLM-5 features three major breakthrough innovations that position it as a competitive domestic model aligned with the Anthropic Opus series [4] - Performance benchmarks indicate that GLM-5 achieved the highest score among open-source models in authoritative programming tests, closely approaching the performance of the leading Claude Opus 4.5, particularly excelling in complex system engineering and long-term project handling [4] - The model's core technology has been significantly upgraded, with a parameter scale reaching 744 billion, and it innovatively employs an asynchronous reinforcement learning framework called "Slime" along with a sparse attention mechanism, enhancing intelligence while substantially reducing deployment costs [4] Group 2: Development Workflow Transformation - Accompanying the model is the ZCode tool, which allows developers to describe requirements in natural language, enabling the model to automatically decompose tasks and direct multiple agents to concurrently complete coding, debugging, and previewing processes [4] - The ZCode tool also supports remote command of desktop agents via mobile devices, significantly extending productivity [4]