Cloudflare's new Clef decision model answers typed questions with probabilities instead of generated text. Batch up to 64 noul, choice and score questions about the same state into a single request, then decide in your own code where each threshold sits.
Cloudflare's Clef is a 27B multimodal decision model on Workers AI, under the model ID @cf/cloudflare/clef. It is not a chat model. You send it a state, such as a support ticket, a log excerpt or an agent's pending action, plus a schema of typed questions, and it returns answers keyed by the question ids you supplied. A noul question returns yes or no probabilities rather than generated text, a choice question takes a map of criteria you define, and a score question takes an ordered list of levels, such as no impact through critical.
The habit most teams carry over from chat models is one question per call. With Clef that wastes money and latency. The model accepts 1 to 64 typed questions per request, so a single call can ask whether a ticket is urgent, which team should own it, how severe the customer impact is and whether it mentions a refund, all against the same state. Cloudflare lists Clef at $0.24 per million input tokens, and the state is the expensive part, so paying for it once instead of four times is the whole game.
The second habit to drop is letting the model decide. Clef gives you probabilities, which means the cutoff is yours. Route to a human when the urgent probability crosses a level you chose after looking at real traffic, not when a generated sentence happens to say yes. Thresholds live in your code, are easy to audit and can be tuned per workflow without touching a prompt.
Watch the limits. The context window is 65,536 tokens, and long text state is truncated to fit the model's token limit, so put the decisive material at the top of the state rather than trusting the end of a long transcript to survive. Clef accepts up to 4 embedded PNG, JPEG or WebP images per request and does not accept remote image URLs, so fetch and embed screenshots yourself.
Pick the size by the job. Cloudflare reports a median latency of 38.8 milliseconds for Clef-flash and 209.3 milliseconds for Clef. Put Clef-flash in the hot path of an agent loop where every step waits on a routing decision, and reserve Clef for decisions where the extra latency buys accuracy you have measured. Clef-flash weights are published on Hugging Face under the Apache 2.0 license if you need to run it yourself.
Try it today
- Call @cf/cloudflare/clef (or clef-flash) with one state and a questions map keyed by your own ids.
- Batch every decision about that state into the same request, up to 64 questions of type noul, choice or score.
- Put the most decisive text at the top of the state, since anything past 65,536 tokens is truncated.
- Embed images directly (PNG, JPEG or WebP, at most 4); remote URLs are not accepted.
- Set and log your own probability thresholds per workflow, and start agent hot paths on Clef-flash.