By putting a screenshot or design mockup and a clearly specified output format into the same request, GLM-5V-Turbo can be used directly for UI recreation, visual localization, and the initial perception step of a GUI or coding agent.
Suitable tasks: Design-to-frontend-page conversion, locating targets in screenshots, visual understanding of documents and pages, and coding tasks performed after observing a video or GUI.
Unsuitable tasks: Complex text-only backend architecture decisions, or tasks that require the model to claim it completed GUI operations without real tool results.
Supported model version: glm-5v-turbo; the official page lists video, image, text, and file input; text output; a 200K context window; and a 128K maximum output.
Supported client, agent, or API: Z.AI Chat Completions API; it can be connected to agents such as Claude Code and OpenClaw, but tool permissions and the execution layer require separate configuration.
Recommended reasoning tier and parameters: The official Quick Start uses thinking.type="enabled"; whether to enable it should be evaluated based on latency and task complexity. For image localization, start with a simple request; for long-horizon recreation, compare it with thinking enabled.
Visual localization (complete example prompt from the official documentation):
Where is the second bottle of beer from the right on the table?
Provide coordinates in [[xmin,ymin,xmax,ymax]] format.Design mockups to mobile pages (official example prompt, suitable for supplying multiple design images in the same request):
Please recreate the mobile pages based on the design mockups in the images.
The left side shows the welcome page, and the center shows the homepage image.
You will also need to create mockups for the remaining two pages.Reusable enhanced working prompt (preserves the official task intent and adds acceptance-check fields):
Recreate the interface shown in the attached design mockups.
Requirements:
- identify every visible page, component, text block, spacing relationship, and interaction affordance
- implement the pages as runnable frontend code
- preserve the visual hierarchy, layout, colors, typography, and responsive behavior
- list any element that cannot be verified from the images instead of inventing it
Output:
1. implementation
2. page/component mapping
3. unverifiable assumptions
4. a short visual QA checklistSend a single screenshot or design mockup and a clear objective first; if there are multiple pages, label each image with the page name or sequence.
For localization tasks, specify the coordinate format; for recreation tasks, specify the run method, page list, and output for “uncertain items.”
Use model="glm-5v-turbo" in API requests; for complex visual reasoning tasks, compare thinking.type="enabled" with disabled or default behavior.
Have the model produce the page or localization result first, then let the agent or browser perform the actual action; the execution layer should return screenshots, errors, and DOM/log evidence.
Run local startup, page screenshots, and visual checks on the frontend result; list anything the model could not confirm from the images as an assumption, not a fact.
The official model is positioned as a “multimodal coding foundation model,” supports understanding images, video, and text, and is intended for long-horizon planning, complex coding, and action execution.
Official capabilities include thinking mode, vision comprehension, streaming, function call, and context caching.
The official use cases list Frontend Recreation, GUI Autonomous Exploration and Recreation, Code Debugging, and OpenClaw.
The official Quick Start uses an OpenAI-compatible chat/completions request whose message content consists of image_url and text, and provides the localization and design-recreation prompts above.
The official page shows example inputs and capability positioning, but does not provide complete generated code, visual scores, or success rates for either example.
Coordinates require confirming the image coordinate system and resolution first; scaling, cropping, and video frames across clients can change what the coordinates mean.
A design mockup recreation prompt cannot fill in business logic, interaction states, or backend data that were not provided; require the model to list assumptions explicitly.
GLM-5V-Turbo’s visual perception does not automatically confer browser or shell access; tool calls should be audited by an external agent.
The official example asks “Provide coordinates in [[xmin,ymin,xmax,ymax]] format” and describes the recreation task as “recreate the mobile pages based on the design mockups.”
GLM-5V Turbo