Clef-GGUF

Clef-GGUF

GGUF files of Cloudflare's Clef decision model, which takes image and text input. Served with llama.cpp tools and called through a decision API path. For developers running it on their own machine.

Made by ggml-org 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: Clef-GGUF is a GGUF conversion of the Cloudflare/clef model, made by ggml-org. The model has 27 billion parameters and accepts images and text, so you can run it locally with llama.cpp tools.

The card calls it a decision model. You use it through the /v1/systemone API path, not as a general chat model. The files were converted automatically with ggml-org's convert tool.

Who it suits: Developers who run models locally with llama.cpp and want to try Clef's decision API. The page is short, so expect to read the original model's page for details on what it decides.

Industry | Who uses it | How they use it | Example
Software | Backend developers | Run a decision model behind a local API | Starting a local server and sending it requests
Research | ML engineers | Test a converted model with llama.cpp tooling | Checking the GGUF build against the original
Edge computing | Self-hosting teams | Keep a 27B model on their own hardware | Running it without a cloud service
Tooling | llama.cpp users | Try a model that needs a recent llama.cpp change | Following the card's setup note

Anyone looking for an everyday chat model can skip it, since the card describes a decision model used through one API path.

Where to use it:

• The Llama app, which the card names as a way to run it.
• A terminal, with the llama serve command.
• Windows, macOS or Linux with Python, for downloading the files.

Use case | Who | What they get back
Local decision API | Developers | Answers from a model on their own machine
Format testing | Tinkerers | A GGUF build of the original model to try
Self-hosting | Small teams | A 27B model without an outside service

How to set it up: You need a llama.cpp setup that supports this model and enough memory for a 27B model.

• A llama.cpp build that includes the specific pull request the card names
• Enough memory, since the 4-bit weights are about 13.5 GB by Hugging Face's estimate
• Python and the Hugging Face tool, if you want the files downloaded separately

1. Check that your llama.cpp has the change the card requires.
2. Optionally download the files with the Hugging Face tool.
3. Start the server with the terminal command below.

pip install -U huggingface_hub
hf download ggml-org/Clef-GGUF

How to use it: Start the model from the terminal, pulling it straight from Hugging Face. Then call the decision API path on the running server.

llama serve -hf ggml-org/Clef-GGUF

Tool | What it does
llama serve | Starts a local server that loads the model
-hf option | Pulls the model files from Hugging Face by name
/v1/systemone | The API path where the decision model is used

Example: Start the server with the llama serve command, then send a decision request to the /v1/systemone path. The card does not show the request body, so look at the original Cloudflare model page for the format. The server returns the model's decision.

For newcomers and experts: If you are new, this is a large model packaged for local use, and the card is brief. Run the server command, then read the original model page to learn what questions it answers.

For experts, this is an automatic GGUF conversion of Cloudflare/clef with image and text input. It relies on a llama.cpp pull request named in the card, so a stock build may not load it. The card gives no quant table or benchmark numbers.

Cloudflare publishes the Clef model
ggml-org converts it to GGUF
You start it with llama serve
You call the /v1/systemone path
The model returns its decision

Good to know: The points below cover limits and costs.

• Requirement: a llama.cpp change named in the card must be present.
• License: Apache 2.0.
• Cost to run: no API key or paid service is named. A 27B model needs a lot of memory.
• Data it reaches: it reads the requests you send. The server command downloads the files from Hugging Face.
• Not for: general chat use, since the card describes a decision model used through its API path.

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: 20,883
• Likes: 30
• License: apache-2.0
• Based on: Cloudflare/clef
• Tags: gguf, quantized, decision-model, image-text-to-text, conversational

Links

• Model: https://huggingface.co/ggml-org/Clef-GGUF
• Files: https://huggingface.co/ggml-org/Clef-GGUF/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download ggml-org/Clef-GGUF

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