Which models on inference.eu accept images?

Vision-capable models are marked as such on their model pages and in the catalogue. `mistral-small-3.2-24b-instruct` is one example. Call `GET /v1/models` to see what your key can use, and choose a larger vision model for dense documents or complex charts.

Is the vision API compatible with OpenAI?

Yes. Images go into chat completions messages as `image_url` content parts, exactly as with OpenAI's vision models, and into the Responses API as `input_image` items. The Anthropic Messages format accepts `image` content blocks. Existing code works after you change the base URL, key and model.

Can I send a local image file?

Yes. Read the file, encode it as base64 and send it as a data URL, such as `data:image/png;base64,...`, in the `image_url` field. This also avoids exposing the image on a public URL. Compress large images first to stay within the request size limit.

How many images can I send in one request?

Several, as long as the total fits within the model's context window and the request size limit. Each image uses tokens in proportion to its resolution. The exact limit per request depends on the model and your plan, and we confirm it with you.

Can a vision model read PDFs?

Vision models read images, so you would render each PDF page as an image. For most PDFs it is better to convert them to Markdown with the document conversion tool and send the text. Use vision for pages where layout, charts or handwriting carry the meaning.

Can I get structured data from an image?

Yes. Combine the image with `response_format` and a JSON schema, for example to extract the fields of a receipt or a delivery note. The model returns JSON that follows your schema. Validate important values such as totals before you use them.

Are the images I send stored or used for training?

No. Images are processed in EU data centres to answer the request and are never used to train models. That is important for scans of contracts, invoices and screenshots that contain personal or confidential information.

Why does the model misread numbers in my image?

Small or blurred text loses detail when the image is converted to tokens. Send a higher-resolution or cropped image, use PNG for documents and ask the model to quote the exact text. For critical numbers, convert the document to text or validate the result against another source.

Vision API for images · OpenAI-compatible, processed in the EU

Vision: send images to open models, processed in the EU

Vision models read images as well as text. You can ask them to describe a photo, read a receipt, compare two screenshots, check a form or answer questions about a chart. Images are sent in the same chat request as text, in the OpenAI or Anthropic format, and are processed in European data centres, which matters when they show documents, people or customer data.

import base64, os from openai import OpenAI client = OpenAI(base_url="https://api.inference.eu/v1", api_key=os.environ["INFERENCE_EU_API_KEY"]) with open("receipt.jpg", "rb") as f: image = base64.b64encode(f.read()).decode() reply = client.chat.completions.create( model="mistral-small-3.2-24b-instruct", messages=[{ "role": "user", "content": [ {"type": "text", "text": "List the items, the VAT rate and the total on this receipt."}, {"type": "image_url", "image_url": {"url": f"data:image/jpeg;base64,{image}"}}, ], }], max_tokens=500, ) print(reply.choices[0].message.content)

How to send an image

Choose a vision-capable model, such as mistral-small-3.2-24b-instruct, and make the user message content a list of parts instead of a string. A text part carries your question; an image_url part carries the image, either as a public HTTPS URL or as a data URL with base64 content, as in the example. You can mix several images and text parts in one message.

The answer comes back as normal text in choices[0].message.content. With the Anthropic Messages format, use an image content block with a base64 or url source instead. With the Responses API, use an input_image item.

Image size, detail and tokens

Images are converted to tokens before the model reads them, and larger images produce more tokens. That affects cost, speed and how much of the context window remains for text. Resize images to what the task needs: a photo for a general description can be small, while a dense invoice or a screenshot of a table needs enough resolution for the text to stay legible.

Use JPEG for photos and PNG for screenshots and documents with sharp text. Crop to the relevant area where you can. The detail field on image_url is accepted for compatibility; how it changes processing depends on the model.

Practical patterns

Receipts and invoices: send the image with a JSON schema in response_format to get supplier, date, items and totals as structured data. Screenshots: ask a model to describe a UI bug, read an error dialog or compare a design with the built page. Quality checks: show a product photo and ask whether a label is present and readable.

For multi-page PDFs, it is often better to convert the document to Markdown first with the document conversion tool and send text, using vision only for pages with charts or handwriting. Text is cheaper per page and easier to search, cite and store.

Errors and edge cases

Sending an image to a text-only model returns a 400 error saying the model does not accept images; check the model page or GET /v1/models. Image URLs must be publicly reachable over HTTPS; private links, expiring links and pages that need a login fail, so send base64 instead. Very large files can exceed the request size limit, so compress them first.

Vision models can misread small print, handwriting and numbers, and they may describe what they expect rather than what is there. For anything that matters, such as amounts on an invoice, validate the values and keep a person in the loop. Do not use vision models to identify people.

Vision: send images to open models, processed in the EU

How to send an image

Image size, detail and tokens

Practical patterns

Errors and edge cases

Vision: common questions

Related

Move your AI workloads to EU infrastructure