Image vision
Hand it a screenshot — an error, a design mock, a chart. The image is spliced directly into the current turn, not converted to a text description first.
Which models support it
| Provider | Vision-capable models |
|---|---|
| GLM | glm-4.6v, glm-4.6v-flash |
| Kimi | kimi-k2.7-code, kimi-k2.7-code-highspeed, kimi-k2.6, kimi-k2.5 |
Switch with /model.
How images enter the context
A user message carrying images is expanded at wire time into OpenAI content blocks — text block first, then image_url blocks, with the image inlined as a data: base64 URL.
Images are dropped on non-vision models
The expansion only happens when supportsVision is true. Switch to a model without vision and the attached images are never spliced into the request — it cannot see what you pasted. Use /model to switch to one of the models above first.
Vision vs document parsing
| Vision | Document parsing | |
|---|---|---|
| Path | The current model's vision capability | GLM-OCR, decoupled from the active provider |
| Good for | Screenshots, charts, UI | PDFs, scans, long documents with layout |
| Output | The model looks at the image directly | Converted to markdown, then handled as text |
An error screenshot: vision. A 40-page PDF: document parsing — once it's markdown it's searchable, quotable text, whereas pointing a vision model at 40 pages of images is both expensive and easy to skim past.
Related: Document parsing · Providers