content becomes an array of parts, text blocks and image_url blocks, with one rule that trips people up: every image must be inlined as a base64 data URI. The field is named image_url for OpenAI compatibility, but the inference engine does not fetch remote URLs.
A remote
https://… URL in image_url.url is silently unresolvable at the inference engine and returns an upstream validation error. If your image lives at a URL, download it in your own code and encode the bytes as data:image/<format>;base64,… before sending.Getting good results
- Ask a specific question next to the image, “extract the total” beats “describe this” when you want a field, and pairs well with JSON mode for machine-readable extraction.
- Images are tokenized into the prompt, so they count as input tokens (and base64 inflates the request body by ~33%), send the smallest image that is still legible; long documents usually work better cropped to the relevant region than downscaled whole.
- Vision input combines with the rest of the API surface, streaming, tools, and JSON mode all work on multimodal requests.
GLM-5.3 Flash is the catalog’s vision model today; text-only models reject image parts. Check
input_modalities on GET /models before routing multimodal traffic.