Cloudflare's Clef: Models That Skip the Chat and Just Decide

Cloudflare's Clef: Models That Skip the Chat and Just Decide

Think about what happens inside an AI agent. Before it writes anything, it makes a dozen small calls: is this ticket urgent, which team owns it, is this crawler a bot or a person, does this tool call look safe. Most teams run those calls through a chat model, pay chat-model prices, and wait for chat-model latency to answer a yes-or-no question.

Cloudflare thinks that is wasteful. On October 1, the company released two models called Clef and Clef-flash that do not write anything at all. They read an input state and a set of typed questions, then return an answer with a confidence score. No prose, no pleasantries, just the decision.

The design is deliberate. Clef is built on a 27-billion-parameter Qwen backbone, and Clef-flash uses a 9-billion-parameter one. Both freeze that backbone and train low-rank adapters on top, which kept training cheap and the models fast. A prefill-only architecture scores the possible answers in parallel instead of generating tokens one by one, so a decision comes back in a fraction of the time a chat model would take.

Each model handles a 64,000-token context window and can answer up to 64 questions in a single request. The questions come in three flavors: yes or no, multiple choice, and scored rubrics. One network round trip can classify a whole batch of support tickets, route a queue of tool calls, or score a set of candidates against a rubric.

The published benchmark numbers are striking. On BANKING77, a standard intent-classification test, Clef scores 94.20 macro-F1 against 79.74 for Jev, the decision model that defined this category. Median latency lands at 209.3 milliseconds for Clef and 38.8 milliseconds for Clef-flash, versus 524.1 milliseconds for the baseline. In other words, the smaller model answers roughly 13 times faster than what teams were using.

The part that matters for local runners: both models ship as open weights under Apache 2.0 on Hugging Face. You can run them on your own hardware instead of calling an API, which is exactly what you want for classification work that touches your private data. They are also live on Workers AI as @cf/cloudflare/clef and @cf/cloudflare/clef-flash if you would rather call them as a service.

None of this replaces a chat model. Clef cannot write, reason out loud, or hold a conversation, and that is the point. The decision to use it is a pipeline-design decision: every classification call you move off your big model saves tokens, cuts latency, and shrinks the context your main model has to carry.

Trying it is cheap. The Hugging Face weights work with the same serving stacks you already use for Qwen-class models, so a quick eval is a weekend project: point your ticket classifier at Clef-flash, compare its confidence scores against your current chat-model labels, and measure the latency and token savings yourself. If the accuracy holds on your data, you have just freed up your big model for the work that actually needs language.

If you are sizing hardware for this, the 9B flash model fits comfortably on consumer graphics cards while the 27B version wants more room. Run the numbers through the VRAM calculator to see what your card can hold before you pull the weights.