GLM-4.7

GLM-4.7 is a Mixture of Experts language model built for intelligent agent workloads, with 358 billion total parameters and 32 billion active parameters per token generation. The GLM-4.x series was architected from the ground up to unify reasoning, coding, and autonomous agent capabilities within a single foundation model, making it particularly well-suited for applications that require sustained multi-step task completion. It supports an interleaved thinking mode that allows the model to pause and deliberate before taking actions, which proves valuable in complex agent frameworks where premature outputs can derail entire workflows. The model's architecture enables strong performance across both English and Chinese, while maintaining the parameter efficiency needed for practical deployment in production agent systems.

The GLM-4.7 release represents a meaningful step forward from its predecessor GLM-4.6, with substantial gains across coding, reasoning, and tool-use benchmarks. On SWE-bench, which tests real-world software engineering problem-solving, GLM-4.7 reaches 73.8%, a 5.8 percentage point improvement, while multilingual coding tasks show even larger gains at 66.7% with a 12.9 point jump. Complex reasoning receives a significant boost as well, with the HLE Humanity's Last Exam benchmark showing 42.8% accuracy, up 12.4 points from the previous version. Tool-using capabilities show marked improvement on benchmarks like τ²-Bench and web browsing tasks via BrowseComp, reflecting the model's readiness for agentic deployments. The model has already been adopted as the default inference provider in popular open-source coding agents including Kilo Code and OpenCode, signaling strong community trust in its practical utility for developer-facing applications.