GLM-4.7

An open model for coding, reasoning and agents with turn-level control and preserved thinking.

Context200K

ModalityText

AccessWeights · API

ReasoningHybrid

Where it fits.

Especially relevant to terminal coding agents and Chinese development workflows, with open weights for customization. GLM-4.7 focuses on multilingual agentic coding, terminal work, tool use and complex reasoning. It adds interleaved, preserved and turn-level thinking to keep long coding sessions more consistent.

Core capabilities

200K context for long coding and agent sessions
Interleaved thinking before responses and tool calls
Preserved thinking across turns plus per-turn control
Open weights with vLLM and SGLang serving paths

Best-fit workflows

Terminal coding agents
Chinese software engineering
Tool and search agents

04 / OPERATIONS

From evaluation
to production.

Deployment notes

Configure glm47 tool parsing and the supported reasoning parser. For long-horizon coding on SGLang, preserve thinking explicitly; disable it for lightweight turns, and validate terminal permissions independently of model quality.

Watch before choosing

Features such as preserved thinking require correct client-side state handling; incompatible implementations can erase the benefit. Preserved thinking depends on client and runtime support; an incompatible template can erase the benefit. Official serving examples require GLM-specific tool and reasoning parsers.

Family & versions

GLM-4.7Capability
GLM-4.7 FlashLight 30B-A3B
GLM-4.6VVision

Specifications and positioning checked against provider documentation.