Changelogs: Week of July 27, 2026

Last week, Qwen models on PolarGrid gained the ability to read images alongside text in the same request.

Qwen vision: image input on chat completions

Qwen models on PolarGrid now read images alongside text in the same request. Images attach as OpenAI-compatible image_url content parts on POST /v1/chat/completions, either as a base64 data: URI or a public https URL. The gateway decodes or fetches the image and passes the pixels to the model, which answers in text. A single request can mix both forms and carry text parts in any order. URL fetches are scoped to public addresses, do not follow redirects, and time out at 10 seconds.

The vision-capable model is qwen-3.6-35b-a3b, a 35B mixture-of-experts model with approximately 3B parameters active per token, served at FP8. Limits are 4 images per request and 20 MB per decoded image. Image support is declared per model; a model served without it returns 400 Bad Request for image content, so a blind text answer is never silently returned in its place.

Types ship in the JavaScript SDK from 0.10.0 and the Python SDK from 0.9.0, which includes an image_url_from_path helper for local files.