GLM-4.7

GLM-4.7 is part of the GLM-4.x series, a foundation model family purpose-built for intelligent agents. The model carries 355 billion total parameters with 32 billion active parameters during inference, a sparse mixture-of-experts design that lets it maintain broad capability while focusing computation efficiently. It was architected to unify reasoning, coding, and agentic tool use within a single system, targeting the growing demand for models that can plan, execute multi-step tasks, and interact with real-world tooling rather than just generating text in isolation.

The model represents a substantial step forward from its predecessor GLM-4.6, posting marked gains across coding and reasoning benchmarks that matter for agent workflows. On software engineering tasks, it reaches 73.8% on SWE-bench, a +5.8% improvement, while multilingual coding climbs to 66.7%, up 12.9%. It also demonstrates stronger mathematical and general reasoning, reaching 42.8% on the HLE Humanity's Last Exam benchmark, a +12.4% jump. GLM-4.7 can think before acting, enabling it to handle complex agent loops in frameworks like Claude Code and Cline more reliably. Tool use and web browsing capabilities have been sharpened as well, showing measurable progress on τ²-Bench and BrowseComp, making it a practical choice for developers building autonomous coding assistants or multi-step automation pipelines.