Describe JPEG XL images with Qwen 3.5 9B, hosted locally by Masland Tech LLC. A compact vision model — much more efficient than frontier models in remote data centers.
Describe JPEG XL images with Qwen 3.5 9B, hosted locally by Masland Tech LLC. A compact vision model — much more efficient than frontier models in remote data centers.
This site is invite-only. Enter the access code to describe images.
JXL Vision describes a JPEG XL image with a local vision model. Drop a .jxl file and the text appears as it is written.
The model is Qwen 3.5 9B (Q4_K_M GGUF, with its matching vision projector) running on llama.cpp. Inference is hosted locally by Masland Tech LLC — not a frontier-model API in a remote data center. A 9-billion-parameter model on a single machine is far more efficient in energy, cost, and hardware than the large models those data centers run.
Your image is uploaded, decoded in memory, sent to the local model, and discarded when the request finishes. It is never written to disk. Other visitors cannot see it or the description; those only exist on your connection.
Each browser gets a private HttpOnly session cookie. Uploads need that cookie plus a CSRF header, so another site cannot submit an image as you. If the operator set an access code, you enter it once per session.
JPEG XL is the only format accepted — no PNG or JPEG conversion. The GPU describes one image at a time; extras wait in a short queue. Rate limits apply per session and per IP.
Metrics are anonymous cumulative charts (page visits, images described, tokens, megapixels) — never filenames, pictures, or text. Do not upload images you would not trust this server’s operator to see.