GLM-4.7: An open model built for real development workflows
Z.ai has open-sourced GLM-4.7, the newest member of the GLM family, released just before Christmas. The focus is simple: handle long-running, tool-heavy engineering tasks with consistent behavior and fewer surprises.
GLM-4.7 ranks #6 in WebDev and is the #1 open model.
What's new compared to GLM-4.6
GLM-4.7 leans harder into engineering use. It strengthens coding workflows, complex reasoning, and agent-style execution, making multi-step tasks more stable and consistent across sessions and tools.
Conversation and writing quality also tightened up. Output is clearer and more economical, and the model behaves like a single coherent system rather than a set of disconnected variants.
Built for long task cycles and tool use
As tasks stretch across hours or days, tiny mistakes add up. Tool calls need to be correct, and behavior has to stay steady across iterations. GLM-4.7 was trained and evaluated with that reality in mind.
It supports "think-then-act" patterns in widely used coding agents and frameworks, including Claude Code, Cline, Roo Code, TRAE, and Kilo Code, matching how developers actually work in terminals and editors.
Results from real programming work
Z.ai evaluated GLM-4.7 across 100 real tasks inside a Claude Code-based environment covering frontend, backend, and instruction-following work. It showed higher completion rates and more consistent behavior than GLM-4.6, cutting down on prompt tweaks and retries.
Based on these results, GLM-4.7 is now the default model for the GLM Coding Plan.
Benchmarks and tool-use performance
- BrowseComp: 67.5
- τ²-Bench (interactive tool use): 87.4 - the highest reported score among publicly available open-source models to date
- Programming suites: Performs at or above Claude Sonnet 4.5 on SWE-bench Verified, LiveCodeBench v6, and Terminal Bench 2.0, with clear gains over GLM-4.6
- Code Arena: First among open-source models and top among models developed in China
- WebDev ranking: #6 overall and #1 open model
More predictable, controllable reasoning
GLM-4.7 offers finer control over reasoning depth and consistency across long runs. It keeps its chain of thought steady across multiple interactions and adapts the depth of reasoning based on task complexity, which makes agent behavior more predictable over time.
Better front-end generation and general output
Front-end work benefits from a stronger grasp of visual structure and established design conventions. Layouts tend to arrive with cleaner spacing, clearer hierarchy, and more coherent styling, which means less manual cleanup.
Conversation quality and writing style also improved, broadening use across documentation, specs, and product-facing content alongside engineering tasks.
Ecosystem integration and access
GLM-4.7 is available through the BigModel.cn API and fully integrated into the z.ai stack. Adoption is already underway across developer tools and infrastructure providers.
- Integrations and partners: TRAE, Cerebras, YouWare, Vercel, OpenRouter, CodeBuddy
The net effect: GLM-4.7 is moving from research settings into everyday engineering and product work.
Try it, ship it
- Default model for the GLM Coding Plan
- Try GLM-4.7 in the chat interface
- Download weights on Hugging Face
- Technical blog: training notes and benchmarks
Level up your team
If you're formalizing AI skills for engineering workflows, this helps:
Your membership also unlocks:
