Skip to main content
In training and evaluation, vision models can read images in chat completion messages. The input_modalities field in those model catalogs describes model-level capability; it does not guarantee that a trained checkpoint accepts images in deployment chat. Checkpoint deployments currently accept text only; requests with image parts return HTTP 400. See deployments. Your harness can show the model a screenshot, chart, or photo in a user message or a tool result. Evaluations score those trajectories, and training learns from them like any other turn. For fixed-choice image classification, see image decisions. That endpoint accepts image parts in input, rather than chat messages; check the decision API and installed SDK support before reusing a chat example.

Choose a vision model

A model reads images when its input_modalities includes "image":
client.training.list_models() reports the same field for training models. Vision models include the Qwen 3.5, 3.6, and 3.8 models, Inkling, and the hosted GPT, Claude, Gemini, and Kimi baselines. A request with an image for a text-only model returns 400 Model does not accept image inputs.

Add an image to a message

Build an image part with ImageInput and put it in the message’s content list, next to any text parts:
harness.py
part() returns the OpenAI image part with the image as a base64 data URL:
Harnesses that call the OpenAI client directly can send the same part. The API also accepts the input_image type and a plain string image_url. It ignores detail. Images may appear in user and tool messages. To return a screenshot from a tool, put the image part in the tool message’s content list with its tool_call_id. Images in system, developer, or assistant messages return 400.

Upload large or repeated images

Inline images travel in the request body, which is capped at 32 MiB, about 24 MiB of images after base64 encoding. A conversation also resends earlier images on every turn. For large images, or ones the model sees on many turns, attach the image to the trajectory once and send a part that names it:
The returned part works only in requests for that trajectory. Uploads use the artifact upload flow, so the image is also kept with the trajectory.

Formats and limits

The platform resizes each image to fit the model before the model reads it, keeping its aspect ratio. Qwen models read at most 16,777,216 pixels, with sides rounded to multiples of 32. Inkling reads a long side of at most 4,096 pixels. The resized copy applies EXIF rotation, converts colors to sRGB, and places transparent areas on white. Images that already fit pass through unchanged. Image tokens count as input tokens. Fewer pixels cost fewer tokens, so set the image_max_pixels run option to downscale images further:
image_max_pixels accepts 1 to 16,777,216 and can only lower the model’s own limit. A training run’s checkpoints keep its value in later evaluations and deployments unless you set it again. Retaining the value does not enable image input for checkpoint deployment chat, which currently accepts text only.