GLM-4.6V Guide: Function Calling With Images and 128K Documents

The usual way to put a vision model inside an agent is clumsy. The model looks at an image, writes a text description, and the text gets passed to tools. Anything the description missed is gone. GLM-4.6V's main change is that images can go straight into tool calls and come straight back out of them, so the visual evidence stays in the loop.

Z.ai published GLM-4.6V in December 2025. This guide covers what that change means in practice, the document-reading use case Z.ai leans on, the limitations its own authors list, and how it compares with GLM-4.5V.

Two models in one release

The release has two sizes. GLM-4.6V is a 106B-parameter model aimed at cloud and cluster deployment. GLM-4.6V-Flash is a 9B model aimed at local deployment and low-latency use. Both are MIT-licensed, and FP8 weights are published for the larger one.

On Z.ai's API there are three tiers: GLM-4.6V at $0.30 per million input tokens and $0.90 output, a FlashX tier at $0.04 input and $0.40 output, and Flash, which Z.ai serves free. OpenRouter listed GLM-4.6V at the same $0.30 and $0.90 at the time of writing, with a 131,072-token context and a 32,768-token maximum output.

All three tiers accept text, images, video and files according to Z.ai's documentation, with text output. The context window was trained to 128K tokens.

Function calling with pixels

Z.ai describes GLM-4.6V as integrating native function calling into its vision line for the first time. Two things follow from that:

• Images as tool inputs. A screenshot or a document page can be passed directly as a tool parameter rather than converted to text first.
• Images as tool outputs. When a tool returns a chart, a search result image or a rendered web page, the model interprets it and folds it into its reasoning.

The second point is the more interesting one for agents. Consider a front-end fix loop: the model edits a component, a tool renders the page and returns a screenshot, and the model checks the screenshot against the request. With a text-only model, that render step needs a separate vision call and a lossy description. Here it is one model in one loop. Tool calling explained covers the general mechanics if you are wiring this up for the first time.

Z.ai also describes rebuilding HTML and CSS from UI screenshots and then editing the result with natural-language instructions, and the GLM-V repository documents bounding-box output for grounding tasks.

Documents read as pages, not text

The second headline use is long documents. GLM-4.6V reads pages as images, so text, layout, tables, charts and figures are all visible to it without an OCR step that flattens them. Z.ai claims a single pass can cover roughly 150 pages of complex documents, 200 slides, or an hour of video, and demonstrates it comparing four companies' financial reports into one table.

That is a real simplification for pipelines that currently run OCR, then table extraction, then a text model. It is also where you should be most careful. A comparative table built from 150 pages is exactly the kind of output where a single misread figure is invisible unless you check it. Treat extracted numbers as drafts and verify a sample against the source.

If you are deciding between this and a retrieval pipeline, multimodal models explained covers how images become tokens, which drives both cost and how many pages actually fit.

The weaknesses Z.ai lists

Credit to the authors: the model card lists remaining issues plainly.

• Pure text question answering still has significant room for improvement.
• The model may overthink or repeat itself, especially with complex prompts.
• It may restate its answer at the end of a response.
• Perception has limits, including counting accuracy and identifying specific individuals.

The first point is practical: do not route text-only traffic to GLM-4.6V because it happens to be deployed. Use a text model for text. The second and third points cost output tokens, so cap max_tokens sensibly and consider disabling thinking for simple visual lookups.

GLM-4.6V over GLM-4.5V, and over Flash

Against GLM-4.5V the comparison is lopsided on paper. Z.ai lists GLM-4.5V at $0.60 input and $1.80 output with a 16K output cap. GLM-4.6V is half that rate on Z.ai, with a 128K context and native function calling. Unless you have prompts validated on 4.5V, start here.

Against its own Flash sibling, the question is where it runs. The 9B Flash model is the one you can host on modest hardware or call for free; the 106B model is the one to use when accuracy on dense documents or multi-step visual tool use matters.

A short evaluation plan

1. Take ten real documents with known answers: figures from tables, totals, dates.
2. Ask GLM-4.6V and Flash to extract them, and score exact matches.
3. Build one tool loop that returns a screenshot and check whether the model acts on what it sees.
4. Record output tokens per task to see how much the repetition issue costs you.

Vision-language models compared has a broader checklist if you are weighing GLM-4.6V against vendors outside the GLM family.