Cloudflare released Clef and Clef-Flash on Oct. 1, two open-weight models designed to score an application’s predefined answer options. They are available through Workers AI and on Hugging Face under Apache 2.0.
Cloudflare lists the 27-billion-parameter Clef for high-precision decisions and the 9-billion-parameter Clef-Flash for latency-critical tasks; both have a 64,000-token context window. The models target cases where an app needs a constrained answer, such as whether a support ticket is urgent or which team should handle it.
A caller sends the ticket and its context, along with a schema defining the allowed questions and answers. A request can include up to 64 decision questions, each framed as yes-or-no, a choice among options, or a score against an ordered rubric. The models return probabilities for the allowed answers in one pass, without generating prose for another program to parse.
The Clef model card also lists text, structured JSON, image and video inputs. This makes Clef a possible routing or classification step inside an agent workflow, rather than a general-purpose agent that plans tasks or writes open-ended explanations.
Cloudflare says a Clef variant scored highest on seven of the ten decision benchmarks it tested. The company’s table shows Jev scoring higher on When2Call and BRIGHT, while a model labeled “DiffusionGemma Jev” led on PhishNChips. Those vendor-reported results vary by task; they do not establish how accurately the models will handle a customer’s live workload.
The release gives teams a way to test a model that returns a constrained decision directly, either through Cloudflare’s hosted API or with the published weights. Each application still needs its own checks, thresholds and fallback for uncertain cases: a probability over the allowed options does not guarantee that the selected answer is correct.
