KNOWLEDGE ATLAS (02513) GLM-5 released: Technology fully upgraded, Agent capability reaches the open-source SOTA.

KNOWLEDGE ATLAS (02513) officially released the new flagship large-scale model GLM-5, which can complete complex system engineering and long-range Agent tasks, demonstrating open-source SOTA-level Agentic Engineering capabilities in various daily work scenarios, with actual usage experience benchmarked against Claude Opus 4.5. The AGI industry is currently transitioning from "Vibe Coding" to "Agentic Engineering" paradigm, with model capabilities shifting from simple dialogue, rapid prototyping, to autonomously solving real-world long-range system engineering tasks. GLM-5 is the new generation base model designed for this trend. At the technical level, GLM-5 has achieved numerous breakthroughs: the parameter scale has been expanded to 744 billion, pre-training data increased to 28.5 terabytes; a new asynchronous reinforcement learning framework "Slime" has been built to fully unleash the model's potential; and sparse attention mechanisms have been integrated for the first time, significantly reducing deployment costs while ensuring the effectiveness of long texts. In benchmark testing, GLM-5's programming capabilities align with Claude Opus 4.5, achieving the highest scores of 77.8 and 56.2 in SWE-bench-Verified and Terminal Bench 2.0 respectively, outperforming Gemini 3 Pro. GLM-5 also possesses 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 -Bench (planning and execution in complex multi-tool scenarios).