Clef: Cloudflare's 27B model makes decisions instead of writing text

What it's about

Cloudflare has released a model with 27 billion parameters called Clef (parameters are, roughly speaking, the settings a model learns during training). It's available in the Ollama library. Unlike a chatbot, Clef doesn't write answer text. It receives a "state" and a catalog of typed questions, and returns decisions along with probabilities. According to its description, it's a model that "turns a state and a schema of typed questions into decisions." The page says Clef is fine-tuned from Qwen3.8-27B (fine-tuning means retraining an existing model for a specific task). The page also mentions a smaller, faster variant called "clef-flash".

How Clef works

The state can be text, JSON or an image plus text, such as a support ticket, a receipt or a planned tool call by an AI agent. Each question has a type, instructions and usually criteria. There are three types:

• noul: a yes/no question; the answer is a value between 0 and 1.
• choice: a selection from labels, such as "billing", "technical" or "other", with a probability for each option.
• score: a rating on levels, for example Routine, Soon, Urgent; the result is a numeric value such as 1.5481.

Ollama builds the prompt for the model itself. In the page's example, a ticket ("I was charged twice, please refund the payment") is assessed like this: team "billing" at 98.88%, refund request 0.9916, urgency 1.5481 on a scale of 0 to 2. The usage data shows 332 input and 0 output tokens: the model doesn't generate text, it reads probabilities directly.

Reading confidence correctly

The "confidence" value ranges from 0 to 1 and shows how concentrated the probabilities are. It is explicitly not the probability that the answer is correct. In case of a tie, the answer follows the order of the options, so you should list your preferred option first.

Usage: CLI, API and SDKs

In the command line, --ask sets a yes/no question, --choice sets selection labels and --level sets levels (lowest first). Examples from the page: "Is a refund being requested?" returns "yes 99%". A receipt image returns "restaurant receipt: yes 97%", category "meals" 94% and "contains alcohol" 94%. You can also safeguard agents: a planned send_email call to all customers with "FINAL NOTICE" is checked for harm, with the result "yes 93%". With --format json you get the complete API response.

You can reach Clef through the Ollama CLI, the Python library (version 0.6.3 or later, ollama.systemone), the JavaScript library or TypeSafe's Python SDK. The latter works only with text and JSON, not images, because it uploads images in a format Ollama doesn't accept. For images, use the API or the CLI.

Cloud limits and costs

• Question names: letters, digits, _, . and - only, up to 100 characters.
• Up to 4 images per request, each at most 4 MiB and 16 megapixels, 8 MiB combined.
• Input is limited to 65,536 tokens; longer inputs are rejected, not truncated.
• Billing is per input token. Cloud use requires a paid plan or credit.

The page points to accuracy figures for four business workflows from "Typesafe Evals", but the actual numbers aren't included in the text we have.

What you need to try it

The source gives no hardware requirements. As a rule of thumb, a 27B model with 4-bit quantization needs roughly 20–23 GB of VRAM or system memory (the 4-bit variant alone is already 18–20 GB according to Ollama), plus about 18–22 GB of disk space. That runs on a graphics card with 24 GB or a Mac with enough unified memory, and much more slowly on a CPU. Without suitable hardware, there's the cloud variant clef:cloud. Locally you need Ollama 0.35.1 or newer.

Our take

Clef shows a trend: instead of having a large language model phrase decisions in free-form text, a specialized model returns structured, machine-readable probabilities. That's useful for ticket routing, receipt checks or safety checks for AI agents. It's more deterministic and cheaper than text generation because no output tokens are produced. How good the model really is remains an open question: the benchmark numbers are missing from the text, and the claims come from the vendor and the Ollama page. Confidence is also not a measure of correctness, and the model needs the System One API, so it isn't a drop-in for ordinary chat applications. The license isn't stated in the source.