GLM-4.7
Coding and reasoning model from Z.ai
GLM-4.7 is a language model released by the GLM Team at Z.ai. It focuses on coding agents, tool use and complex reasoning tasks. The model supports interleaved and preserved thinking modes to improve stability in multi-turn interactions.
- Scores 73.8% on SWE-bench Verified
- Supports preserved thinking for multi-turn tasks
- Available via Z.ai API and open weights
- Improves UI and slide generation quality
- Compatible with vLLM and SGLang frameworks
What it is
GLM-4.7 is a large language model developed by the GLM Team at Z.ai. It serves as a coding partner with enhanced capabilities in agentic workflows and terminal-based tasks. The model improves upon its predecessor, GLM-4.6, in areas such as multilingual coding, user interface generation and mathematical reasoning. It is designed to integrate into developer environments and support complex, multi-step problem solving through structured thinking processes.
How it came about
The model follows GLM-4.6, offering significant performance gains in coding and reasoning benchmarks. The GLM Team introduced new thinking mechanisms, including interleaved and preserved thinking, to enhance consistency across long conversations. These features build on capabilities first seen in GLM-4.5. The release includes technical reports and blog posts detailing improvements in agent frameworks and tool usage.
Inside the model
GLM-4.7 uses a decoder-only architecture. It introduces preserved thinking, which retains reasoning blocks across multi-turn conversations to reduce information loss. Interleaved thinking allows the model to reason before every action or tool call. Users can control this behaviour per turn, disabling it for simple requests to save cost or enabling it for complex tasks. The model supports local deployment via vLLM and SGLang frameworks, with specific configurations for speculative decoding.
What it can do
The model shows improved scores on SWE-bench and Terminal Bench compared to GLM-4.6. It achieves 73.8% on SWE-bench Verified and 41% on Terminal Bench 2.0. In reasoning, it scores 42.8% on HLE with tools. It also performs well on BrowseComp for web browsing tasks. The system generates cleaner web interfaces and slides. It supports various agent frameworks like Claude Code and Kilo Code, improving instruction following and generation quality.
How to use it
Weights are available on Hugging Face under an open licence. Developers can run the model locally using vLLM or SGLang with FP8 quantisation support. API services are provided through the Z.ai platform. The model requires specific parameters to enable or disable thinking modes during inference. Documentation provides Docker images and pip installation instructions for local serving.
Questions people ask about GLM-4.7
What is preserved thinking in GLM-4.7?
Preserved thinking allows the model to keep reasoning blocks from previous turns in a conversation. This helps maintain context and consistency for long, complex tasks without re-deriving logic from scratch.
How can I run GLM-4.7 locally?
You can deploy the model using vLLM or SGLang frameworks. The developers provide Docker images and pip installation commands. FP8 quantised versions are supported for efficient inference.
What improvements does GLM-4.7 have over GLM-4.6?
It shows gains in multilingual coding, terminal tasks and tool use. Benchmark scores improved on SWE-bench and HLE. The model also offers better UI generation and more stable agentic behaviour.
Good reading and sources
There is no English Wikipedia article on GLM-4.7 yet, so the facts on this page come from the sources listed above: huggingface.co/zai-org/GLM-4.7. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.