Answers yes/no, multiple-choice and rating questions you write about a photo or text, with odds for every option

You give Clef a photo or a piece of text and a short list of questions. Each question has a fixed shape: a yes or no ("Is this photo safe to publish?"), a pick from options you name ("Which room is this: kitchen, bedroom, living room or bathroom?"), or a rating on a scale you describe ("How cluttered is it, from spotless to very cluttered?"). It answers all of them at once and puts a percentage on every option, so you see not just "living room" but how sure it is and what came second.

The difference from asking a chatbot is that Clef never writes a reply. There is no paragraph to read, nothing to parse, and no invented detail to check. You define the possible answers, so adding a new label or a new question needs no retraining, and a short question set comes back in a fraction of a second.

Cloudflare built it on Alibaba's Qwen 3.8 model, which is why it can see images, and already uses it to sort websites into categories with odds such as 95% fashion and under 1% phishing. It is free to download under the Apache 2.0 license.

The Clef demo Space runs in your browser. Pick the living room image example and press Decide: you get the room type, whether it matches the "minimalist" style the listing claims, and a clutter rating, each with its own percentage bars. Then upload one of your own photos and edit one option's description, say "tidy" to "ready for a listing photo", and watch the percentages shift. It runs on Hugging Face's shared free GPUs, so a free account gets a limited number of runs, and the first one can be slow while the model loads. Running it yourself needs a data center GPU.

Clef judges, it does not think things through. On Cloudflare's own tests it falls well behind general reasoning models on hard multi-step questions, and on full business workflows such as invoice handling it picks the exact right set of actions only about two times in three. Keep questions to things a person could judge at a glance, and check the percentages against your own examples before you trust a threshold.

• Check rental or marketplace photos: right room, matches the description, safe to publish.
• Route support messages by team and urgency, and send low confidence cases to a person.
• Screen user uploaded photos on a review site before they go live.
• Tag a folder of product shots with your own categories.
• Gate an automation: only act when the yes is above 90%.

Try the demo

View model page

Read this issue on neuralexpedition.com →