Zhipu doubles down on open weights with GLM-4.7, using lower costs and customisation to win US developers and challenge proprietary Western AI leaders.
Open source is fast becoming China’s sharpest competitive weapon in artificial intelligence, and Zhipu is following DeepSeek’s playbook to prove it.
The company is positioning open models as a distribution strategy rather than a technical preference: cheaper to run, easier to customise, and harder to be cut off from. That approach is accelerating global reach, with US developers already adopting Chinese coding agents. Roughly seven of the top 10 AI models now come from Chinese labs across open and closed ecosystems, and firms are climbing leaderboards at speed.
Zhipu’s latest push comes through the GLM-4.7 series, its flagship foundation model built for Agentic Coding. The model focuses on enhanced coding, long-term task planning, and tighter tool collaboration, and ranks strongly among open-source systems on public benchmarks. An update, GLM-5, has also been released in Hong Kong.
Technically, GLM-4.7 strengthens multilingual programming, on-device intelligent agents, and a “think before you act” workflow across frameworks such as Claude Code, Kilo Code, TRAE, Cline and Roo Code. It scores 67 on BrowseComp and 84.7 on τ²-Bench for tool use, reportedly surpassing Claude Sonnet 4.5. Reasoning improves to 42.8 percent on HLE, ahead of its predecessor and reportedly beyond GPT-5.1, while dialogue and front-end generation are more natural and polished.
Adoption is rising fast. “Our online traffic has increased five-fold. This current subscription limit is a protective measure we have implemented to ensure a stable and high-quality experience for our existing users,” said Zhipu’s GLM coding agent team.
After DeepSeek’s breakout, investors and rivals, including NVIDIA, are watching closely as open source becomes China’s fastest route to AI leadership.
