GLM-4.5V Guide: Grounding, GUI Agents and a 16K Output Cap

Most vision-language models are judged on whether they can describe a picture. That is the least interesting thing GLM-4.5V does. Its more useful trick is telling you where something is in an image, as coordinates you can act on, which is the capability GUI agents and document pipelines actually need.

Z.ai released GLM-4.5V on 11 August 2025. It has since been followed by GLM-4.6V, and for most new work that successor is the better choice. But GLM-4.5V is still listed on Z.ai's API and on OpenRouter, and understanding its shape tells you what to check in any vision model.

Built on Air, not bolted on

GLM-4.5V is built on GLM-4.5 Air, the 106B-total, 12B-active text model, and continues the approach of Z.ai's earlier GLM-4.1V-Thinking research. Z.ai's documentation gives the same 106B total and 12B active figures for the vision model. The weights are MIT-licensed.

Because the language side is Air, the vision model inherits Air's hybrid reasoning: a thinking mode for harder visual problems and a quick-answer mode for simple ones. On Z.ai's API you enable it with "thinking": {"type": "enabled"}; with vLLM or SGLang it is on by default and enable_thinking: False switches it off.

Z.ai claims state-of-the-art results among models of its scale on 42 public vision-language benchmarks. That is the vendor's selection and its definition of scale, and it compares GLM-4.5V with mid-2025 peers. Use it as a reason to test, not as a ranking.

Grounding: the feature to actually use

Ask GLM-4.5V to locate an element and it returns a bounding box. The model card documents the format precisely, and you need to know it before you parse anything:

• The box is wrapped in the special tokens <|begin_of_box|> and <|end_of_box|>.
• It is four numbers, [x1, y1, x2, y2], for the top-left and bottom-right corners.
• The values are relative coordinates on a 0 to 1000 scale, not pixels.
• The card notes that bracket style can vary, so parse tolerantly.

To click on a 1920 by 1080 screenshot, you scale: pixel x is x / 1000 * 1920, pixel y is y / 1000 * 1080. Forget that step and the raw values get treated as pixels, so every click drifts up and to the left of its target, which is a common first bug.

This is what makes the GUI-agent use case work. The model card lists screen reading, icon recognition and desktop operation assistance among the model's GUI tasks, and Z.ai open-sourced a desktop assistant app for debugging alongside the release. A loop of screenshot, locate, act, screenshot again is feasible with a model that returns coordinates reliably. Vision-language models compared covers what else to test before building that loop.

The output ceiling you will hit

Z.ai documents a 16K maximum output for GLM-4.5V, and OpenRouter lists 16,384 tokens with a 65,536-token context. Z.ai's own docs do not state the context figure, so treat the 64K as the catalogue's value.

For visual question answering, 16K output is plenty. For two jobs Z.ai advertises, it is tight:

• Screenshot to HTML. A full page with styles can run long, and thinking tokens count against the same budget.
• Long document analysis. Summarising one page is fine. Translating a dozens-of-pages document in one call is not.

The fix is to chunk: one page or one component per call. If you find yourself chunking constantly, that is the signal to look at GLM-4.6V instead.

Video is also on the capability list in Z.ai's docs. OpenRouter's catalogue lists only text and image input for this endpoint, so if video matters to you, confirm what your provider actually accepts.

GLM-4.5V or GLM-4.6V

This decision is less balanced than most. GLM-4.6V is the same size class, but Z.ai lists a 128K context, native function calling with images as tool inputs, and a lower price: $0.30 input and $0.90 output per million tokens on Z.ai's pricing page, against $0.60 and $1.80 for GLM-4.5V.

Half the rate with a larger context and better tool support means GLM-4.5V is mainly worth keeping where you have already built and validated on it. Grounding prompts and coordinate parsing tuned to one model do not always transfer cleanly, so re-test before swapping.

One caveat on tool calling: OpenRouter lists tool calling for GLM-4.5V, but Z.ai describes native function calling as arriving for the first time with GLM-4.6V. If your agent depends on structured tool calls with image inputs, test that path on 4.5V specifically rather than assuming it.

How to evaluate it on your own screens

1. Collect thirty screenshots from the application you actually want to automate, including awkward ones: dark mode, dense tables, small icons.
2. For each, write the element you want located and its true box.
3. Measure whether the predicted box centre falls inside the true box. That is the number that predicts whether a click works.
4. Run it with thinking on and off. Simple locate tasks often do not need it.

Run the same set on GLM-4.6V. If both pass, the cheaper one wins; if neither does, no benchmark score was going to save you.