TL;DR: Cloudflare Clef is a pair of open-source decision models released on October 1, 2026: Clef (27B) and Clef-flash (9B). Instead of writing text, they take a “state” (text, JSON or images) plus up to 64 typed questions and return probabilities for each allowed answer. On Workers AI, Clef costs $0.24 and Clef-flash $0.038 per million input tokens. The weights are on Hugging Face under Apache 2.0, and llama.cpp v0.6.0 can serve them locally through its new /v1/systemone endpoint. Use one when your code needs a fast yes/no, a category or a score, not prose.
Last updated: October 23, 2026. Checked against Cloudflare’s launch blog and Workers AI docs, the Hugging Face model cards and llama.cpp’s server docs.
Most AI tools generate text. A growing group of small, fast models does something narrower: it makes a decision. Cloudflare’s Clef is the most prominent open-source example so far, and llama.cpp added a dedicated endpoint for this model type the same week.
This explainer covers what decision models are, what Cloudflare Clef can and can’t answer, what it costs, and how to try one on your own machine. We haven’t run Clef on our own workloads, so the benchmark figures below are Cloudflare’s, and we say so where we quote them.
What is a decision model?
A decision model answers a fixed set of typed questions about an input and returns probabilities, not free text. Cloudflare’s launch post describes it as a model “that produces bounded structured outputs cheaply, quickly and consistently” and can be dropped into a workflow “when a decision is required.”
Its example: pass in a customer support message, ask whether it’s urgent and which team should handle it, and get back typed answers with probabilities. Your code then routes the ticket, escalates it, or hands it to a person.
Decision model vs LLM
| Decision model (e.g. Clef) | General LLM | |
|---|---|---|
| Output | Probabilities for allowed answers | Free text, tool calls |
| Output tokens | Zero (llama.cpp reports output_tokens: 0) |
Many |
| Speed | Single prefill pass, then scoring | Token-by-token generation |
| Consistency | Bounded to your schema | Can drift from instructions |
| Good for | Routing, triage, tagging, moderation checks | Writing, reasoning, agents |
Cloudflare says Clef runs its Qwen backbone in a “prefill-only pass, then scores the valid schema choices in parallel,” with no intermediate text generated. That’s why it’s fast. The trade-off is that it can’t explain itself or do anything outside the questions you ask.
What can Cloudflare Clef answer?
Three question types, up to 64 per request: yes/no, choice and score. Per the Workers AI model page:
noul(yes/no): returns the probability the answer is true, e.g. “Is this request urgent?”choice: picks from named options and returns a probability for each, e.g. billing, technical or sales.score: rates on an ordered scale you define, e.g. “No impact” to “Critical,” and returns a probability-weighted score.
Each request takes 1 to 64 questions. The state can be a string or structured data such as records, chat logs or app state. Clef also accepts up to four images (PNG, JPEG or WebP, 4 MiB each) as an extension to the API. llama.cpp’s docs add that Clef “reads all the questions in one prompt and decides them jointly,” while other decision models answer each question independently, and that a choice question can have up to 255 options with Clef.
Clef vs Clef-flash: 27B vs 9B
Clef is the more accurate 27B model; Clef-flash is a 9B model built for low latency and a lower price. Both are multimodal and have a 65,536-token context window, per Cloudflare’s docs.
| Clef | Clef-flash | |
|---|---|---|
| Size | 27B (Qwen3.8-27B backbone) | 9B (Qwen3.5-9B backbone) |
| Workers AI price | $0.24 per M input tokens | $0.038 per M input tokens |
| Context | 65,536 tokens | 65,536 tokens |
| Median latency (Cloudflare’s tests) | 209.3 ms | 38.8 ms |
| Vision | Yes | Yes |
| License | Apache 2.0 | Apache 2.0 |
Sources: Clef model page, Clef-flash model page, Cloudflare blog.
On Cloudflare’s own benchmark table, the two trade wins. Clef-flash scored higher on API-Bank accuracy (93.11 vs 91.93) and a home appliances task (97.73 vs 82.95), while Clef was far ahead on CLINC150+OOS intent detection (97.43 vs 66.77 macro-F1). These are vendor numbers, published at launch, and not independently verified. Our reading: start with Clef-flash, and move up to Clef if accuracy on your own labeled examples isn’t good enough.
How much does Cloudflare Clef cost?
On Workers AI, Clef is $0.24 and Clef-flash $0.038 per million input tokens, and there are no output tokens to pay for. Workers AI bills in “Neurons” behind the scenes at $0.011 per 1,000, and every account gets 10,000 Neurons per day free. Usage beyond that needs the Workers Paid plan.
Running the open weights yourself costs nothing beyond your hardware. Cloudflare also says it doesn’t read, store or train on Workers AI requests or responses unless you opt into its fine-tuning service.
How do I call Clef on Workers AI?
Send a state and a questions object to @cf/cloudflare/clef, set the model to “clef” or “clef-flash”. From a Worker, that’s env.AI.run("@cf/cloudflare/clef", {...}). From anywhere else, POST to the REST endpoint with an API token:
curl https://api.cloudflare.com/client/v4/accounts/$CLOUDFLARE_ACCOUNT_ID/ai/run/@cf/cloudflare/clef \
-X POST -H "Authorization: Bearer $CLOUDFLARE_AUTH_TOKEN" \
-d '{
"model": "clef-flash",
"state": "Checkout has been failing for every customer for the last hour.",
"questions": {
"urgent": { "type": "noul", "instructions": "Is this support request urgent?" },
"severity": { "type": "score", "instructions": "How severe is the impact?",
"criteria": ["No impact", "Minor", "Major", "Critical"] }
}
}'
The answers come back under the same ids you used: a probability for urgent, and a weighted score from 0 (lowest level) for severity. This example is adapted from Cloudflare’s docs.
How to run a decision model locally with llama.cpp /v1/systemone
Install llama.cpp v0.6.0 or later, serve a Clef GGUF, and POST to /v1/systemone. The v0.6.0 release (October 5, 2026) added full Clef support, text and vision, and the new server API. The ggml-org/Clef-Flash-GGUF model card gives the command:
- Install llama.cpp (
curl -LsSf https://llama.app/install.sh | sh, orirm https://llama.app/install.ps1 | iexon Windows). - Run
llama serve -hf ggml-org/Clef-Flash-GGUF. - Send questions to
http://127.0.0.1:8080/v1/systemonewith the samestateandquestionsshape shown above.
A few things the server docs flag. A server running Clef only serves this endpoint, with no text generation. The whole prompt must fit in --ubatch-size. Image input needs the model’s multimodal projector (--mmproj). Streaming isn’t supported. The docs also warn that probabilities “are not guaranteed to be calibrated for your data,” so test against examples you’ve already labeled before trusting a threshold.
The same endpoint serves other decision models llama.cpp supports, including laya, julia-1, lev, openjev, kev and nimble. Ollama also added Clef and Clef Flash through /v1/systemone on port 11434.
Is Clef free and open source?
Yes. Both models are on Hugging Face under the Apache 2.0 license. The official repos are Cloudflare/clef and Cloudflare/clef-flash, with community GGUF and MLX conversions alongside. Apache 2.0 allows commercial use and modification. The hosted Workers AI version is paid beyond the free daily allocation.
Open weights also mean you can pin a version. Hosted models get retired (see our note on GPT-5.5’s retirement), but a file on your disk doesn’t change.
When should you use a decision model?
Use one when the output is a label, a yes/no or a score that code acts on immediately. Good fits include support ticket routing, tagging content, moderation pre-checks, picking which tool an agent should call, and flagging suspicious domains, which is how Cloudflare says its threat intelligence team tested Clef.
Skip it when you need an explanation, a summary or open-ended reasoning; an LLM is the right tool there. Many setups use both: a decision model on the hot path to choose, and an LLM only when there’s something to write.
Verdict
Cloudflare Clef is the easiest way to try decision models today. Clef-flash on Workers AI costs about 1.1 cents per thousand 300-token requests at $0.038 per million input tokens, and the same Apache-licensed weights run locally through llama.cpp’s /v1/systemone. Start small: pick one routing or tagging decision you already make by hand, label 50 to 100 examples, and see whether Clef-flash’s probabilities line up before you wire it into production.
FAQ
What is Cloudflare Clef?
Clef is a family of open-source decision models from Cloudflare, released October 1, 2026. They return probabilities for typed questions (yes/no, choice, score) instead of generating text.
What’s the difference between Clef and Clef-flash?
Clef is a 27B model aimed at accuracy; Clef-flash is a 9B model aimed at speed. On Workers AI they cost $0.24 and $0.038 per million input tokens, and both have a 65,536-token context.
How is a decision model different from an LLM?
An LLM generates text token by token. A decision model scores a fixed set of answers in one pass and produces no output tokens, which makes it faster and keeps its answers within your schema.
Is Cloudflare Clef open source?
Yes. Clef and Clef-flash are on Hugging Face under the Apache 2.0 license, which permits commercial use.
Can I run Clef locally?
Yes. llama.cpp v0.6.0 or later can run llama serve -hf ggml-org/Clef-Flash-GGUF and answer requests at /v1/systemone. Ollama supports Clef through the same API.
How many questions can Clef answer per request?
Up to 64 per request on Workers AI. With Clef, a choice question can have up to 255 options, according to llama.cpp’s docs.
Sources checked October 6, 2026. Benchmarks are Cloudflare’s own and haven’t been independently reproduced by TrustLoud.
