Cloudflare Open-Sources Clef: AI Models That Don't Write — They Just Decide

On October 1, Cloudflare released two new models, Clef and Clef-flash — the first models the Workers AI team has trained itself. Their defining feature is counterintuitive: they generate no text at all — they only output probabilities.

Cloudflare calls them “decision models.” They work completely differently from chat models. You feed the model an input state plus a set of questions you define yourself, and it returns a probability for each allowed answer. Your code can branch directly on the numbers without parsing a word of text.

Ask the Way You Want, Get Answers the Way You Asked

Clef supports three question types:

• noul: yes/no questions, returning the probability of “yes.” For example: “Is this ticket urgent?”
• choice: multiple choice, returning the probability for each option plus a confidence score. For example: “Which team should take this?” with billing, tech, or sales as options.
• score: scores on a scale you provide, returning a probability-weighted score. For example: rating user-submitted content against your own risk scale.

A single request can carry up to 64 questions plus 4 images. The vision encoder is something Clef has and its rival doesn’t: Typesafe’s Jev reads text only. The example is telling: a support ticket comes in, and one call answers “urgent, who takes it, how severe” — routing, blocking, and human handoff afterward are all plain program logic.

The Numbers: Latency and Accuracy Both Look Strong

Cloudflare’s pitch is speed. Its models run on Cloudflare’s edge GPUs, close to users, so network round-trips are short. Median latency across Cloudflare’s own 43-test benchmark: 209.3 ms for Clef, 38.8 ms for Clef-flash, and 524.1 ms for Jev. At p95, Clef-flash comes in at 122.4 ms versus 536.0 ms for Jev. Note these are Cloudflare’s own test numbers — no third party has reproduced them yet.

On accuracy, Clef took first place in 7 of 10 decision benchmarks. On the BANKING77 classification task, Clef’s macro-F1 is 94.20 against Jev’s 79.74. On the BFCL tool-calling benchmark, Clef-flash’s case-exact score reached 98.76. In Typesafe’s own workflow benchmark, Clef also beat Jev across three scenarios: invoice processing, customer support, and security incidents. The Register notes that the training data hasn’t been disclosed.

The lineup is clear: Clef is the 27B-parameter accuracy model, Clef-flash the 9B-parameter latency model, both with a 64K-token context window. Weights are on Hugging Face under the Apache 2.0 license and can be self-hosted. Hosted pricing on Workers AI: $0.24 per million input tokens for Clef, $0.09 for Clef-flash, with output uncharged.

Compatible with Jev — Just Swap the Endpoint

Clef speaks the same API format as Jev, which Cloudflare calls the “System One API.” Projects already on Jev can switch by changing the endpoint and model name. In other words, Cloudflare is aiming directly at Jev’s user base.

Cloudflare also launched an RL fine-tuning service alongside the models: you can fine-tune Clef on your own request data and redeploy it. It starts as one-on-one collaboration with Cloudflare’s engineering team, with a self-serve version to follow that lets you fine-tune and redeploy on Cloudflare’s infrastructure using your own data. Cloudflare promises that unless you opt into the fine-tuning service, it will not read, store, or use your request data for training.

A New Lane

In the span of three weeks, this is the third public decision model: Typesafe launched Jev first on September 15, OpenAI opened limited preview of its Decisions API on September 29, and Cloudflare followed on October 1. The same day, Amazon put out its own open-source Strands Decider 2B. Four players colliding in one week signals a direction validated from multiple sides.

The use cases are concrete. Agent guardrails: before an agent calls a tool, spend a few dozen milliseconds asking the model “should this step happen?” — cheap and auditable. Ticket triage, trust-and-safety scoring, threat-intel classification — Cloudflare’s own example: a domain-classification task with Browser Run finished in 2.2 seconds with Clef, versus 4.7 seconds with the general-purpose gpt-oss-120b, with fewer labels to boot.

At its core, this splits the LLM’s “slow thinking” into “fast decisions.” Keep calling the big model when you need reasoning; when you need a number to make a call, use a small, cheap, fast model like Clef. Now that agents are moving into production, “can we make countless small judgments quickly and reliably” has become a hard requirement — that’s why this wave of decision models shipped all at once.