Cloudflare clef GGUF

Cloudflare clef GGUF

Static GGUF quantizations of Cloudflare's clef decision model, which scores every answer option to typed questions about a state in one pass instead of writing text. For developers who run it with llama.cpp.

Made by bartowski on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: This repository offers GGUF quantizations of Cloudflare's clef, made with llama.cpp. Clef makes decisions rather than prose. You give it a situation as text or JSON plus a list of typed questions. In one pass it returns a probability for each allowed answer to each question. Because it never writes text, you have no chat, no completions and no prompt template to set up.

Who it suits: It suits developers who want fast routing and triage decisions with probabilities they can act on, and who can run llama.cpp's server.

Industry | Who uses it | How they use it | Example
Customer support | Support teams | Route tickets, flag anger and rate urgency in one call | Scoring a message about a double charge
Operations | Ops teams | Sort incoming requests into the right queue | Choosing between billing, shipping and technical
Software | Developers | Replace prompt-and-parse steps with typed questions | Getting scores for several questions at once
Automation | Workflow builders | Gate a step on a returned probability | Escalating when urgency scores high

If you want a model that chats or writes text, you can skip it.

Where to use it: You serve it with llama.cpp's llama-server and call its system-one endpoint.

• llama-server from llama.cpp, which the card says serves the endpoint
• Text input, or images when you add an mmproj file
• Windows, macOS or Linux with Python, to download the files

Use case | Who | What they get back
Ticket routing | Support teams | A probability for each team option
Sentiment flag | Operations | A probability that a customer is angry
Urgency rating | Developers | A probability for each urgency level

How to set it up: You need llama.cpp and one quantized file. The card made these files with llama.cpp build b11430. The repository holds every quant, so the card points to a download section for fetching just one.

• llama.cpp with llama-server
• One GGUF file, with Q4_K_M as the card's default pick
• Enough disk space and memory for the size you choose

1. Choose a file from the table, Q4_K_M or larger.
2. Download it from the repository.
3. Start llama-server with that file.

Title: Files named in the card
Quant | Size | Use
bf16 | 54.06 GB | Full weights, stored in split parts
Q8_0 | 28.80 GB | Top quality, more than most people need
Q6_K_L | 23.96 GB | Near-perfect quality, recommended
Q6_K | 23.34 GB | Near-perfect quality, recommended
Q5_K_M | 20.64 GB | High quality, recommended
Q4_K_M | 17.65 GB | Good balance, the card's default pick

pip install -U huggingface_hub
hf download bartowski/Cloudflare_clef-GGUF

How to use it: Send a request to the server's system-one endpoint with a state and a set of questions. The llama-server README documents the exact request and reply formats. The card's example uses three question types.

Tool | What it does
llama-server | Serves the model from a GGUF file
systemone endpoint | Takes a state plus questions and returns probabilities
choice question | Picks among named options, such as billing, shipping or technical
score question | Rates along ordered levels, from can wait to right now
noul question | Asks one question, such as whether the customer is angry

Example: You send a customer message saying they were charged twice last week and nobody replied. You add three questions: which team should handle it, whether the customer is angry, and how urgent it is. The server returns a probability for every option of every question.

For newcomers and experts: A decision model answers multiple-choice style questions with numbers, not sentences. Start by downloading Q4_K_M and sending one request with a single choice question.

For experts, the decision head, which turns final hidden states into answers, stays at Q8_0 in every quant. No imatrix exists for this model, because llama.cpp's imatrix tool watches token probabilities from generated text. So every file is a static quant. Perplexity and KL divergence cannot be measured either.

You send a state and typed questions
llama-server runs a single forward pass
The decision head scores each option
Probabilities come back for every question
Your app acts on the numbers

Good to know: The files under 4 bits per weight, meaning the Q3 types, IQ3_M and Q2_K, give up noticeably more quality than an imatrix build of that size would. Formats that need an imatrix, such as IQ1 and IQ2, are not offered. The file descriptions are general guidance, not measurements of this model.

Title: File size by quant type
bf16: 54.06 GB
Q8_0: 28.80 GB
Q6_K_L: 23.96 GB
Q6_K: 23.34 GB
Q5_K_M: 20.64 GB
Q4_K_M: 17.65 GB

• License: Apache 2.0
• Cost to run: none named
• Data it reaches: the state you send, as text or JSON, and images if you add an mmproj file
• Not for: chat or text generation, since it has no completion endpoint and returns only probabilities

On Hugging Face, 13 November 2026

• Task: image text to text
• Parameters: 27B
• Weights: about 13.5 GB at 4-bit (estimate)
• Formats: GGUF
• Downloads in the last 30 days: 18,354
• Likes: 36
• License: apache-2.0
• Based on: Cloudflare/clef
• Tags: gguf, clef, cloudflare, systemone, qwen3.8, post-train, image-text-to-typed-output, multimodal, structured-output, classification, custom-code, image-text-to-text

Links

• Model: https://huggingface.co/bartowski/Cloudflare_clef-GGUF
• Files: https://huggingface.co/bartowski/Cloudflare_clef-GGUF/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download bartowski/Cloudflare_clef-GGUF

Then follow the usage in its model card. Nothing in this kit runs by itself.