Clef (System One)

• 🧠 27B multimodal decision model from Cloudflare, not a text generator
• 🎯 Returns calibrated probabilities for every option of every typed question
• ⚡ Scores a whole schema in a single forward pass
• 👁️ Reads state as text, JSON, images, or video
• 📏 Cloudflare documents a 64k context window; this catalog entry lists 32,768
• 🆕 Adds a vision encoder over text-only predecessor Jev
• 🔧 Open-weight release, served via Workers AI and local runtimes
• 🏢 Trained by Cloudflare as part of its decision-model family

Cloudflare

Cloudflare is an internet infrastructure and developer platform company, reachable at cloudflare.com, that describes its mission as powering the next generation of applications. Alongside its core platform work, it has begun publishing models aimed at…

Clef is Cloudflare's open-weight "decision model": instead of generating prose, it takes a state plus a schema of typed questions and returns a probability for every allowed option of every question. Cloudflare frames this as infrastructure for agents — your code receives typed answers with probabilities and can route a ticket, escalate, or defer to a human without a text-parsing step in between. The model is a 27B multimodal system that accepts the state as text, JSON, images, or video.

The direct same-family predecessor is Jev, Cloudflare's earlier decision model. Cloudflare describes two concrete generational changes: Clef adds a vision encoder, so it can classify visual content, whereas Jev handles text classification only; and Clef's context window is 64k tokens versus Jev's 32k, letting callers pack more state into a single decision. Note that this catalog entry records a 32,768-token window, which may reflect a specific endpoint's limit rather than the model card figure.

Architecturally, Clef and the smaller Clef-flash derive schema choices directly from internal backbone representations rather than emitting intermediate text, using a two-stage attention routing process in which each valid choice extracts prompt-relevant context and fields cross-attend to one another and back to the payload before scoring. Cloudflare publishes internal results on its Decision Index suite and on Typesafe Evals business workflows alongside latency measurements on the model card. A wider-input sibling, Clef-omni, extends the same formulation to audio on a mixture-of-experts backbone.