Z.ai has announced the release of GLM-4.7, the latest iteration of its large language model. According to the company's technical report, this update focuses on enhancing capabilities in coding, tool usage, and complex reasoning compared to its predecessor, GLM-4.6.

The announcement highlights specific benchmark improvements, including a 5.8 percentage point increase on the SWE-bench coding benchmark (to 73.8%) and a 12.4 point gain on the Humanity’s Last Exam (HLE) reasoning benchmark with tool use (to 42.8%). The company also notes advancements in "vibe coding," referring to the model's ability to generate cleaner webpage layouts and better-formatted slides.

Reported Benchmark Performance

A provided comparison table places GLM-4.7's results alongside other leading models like GPT-5, Claude Sonnet 4.5, and DeepSeek-V3.2 across 17 benchmarks for reasoning, coding, and agent tasks. In this comparison, GLM-4.7's results are competitive, leading in some categories like SWE-bench Multilingual while trailing behind others like GPT-5.1 High on several reasoning benchmarks.

Integration and Availability

GLM-4.7 is now accessible through Z.ai's platform and API. The model weights are also publicly available on Hugging Face and ModelScope for local deployment, supporting inference frameworks like vLLM. For developers using coding agents like Claude Code or Cline, the model is available as an option for subscribers to Z.ai's GLM Coding Plan.

The company promotes this plan as offering access to a high-level coding model at a lower cost than some alternatives, with subscribers being automatically upgraded to GLM-4.7.