clef-flash-4bit

clef-flash-4bit

A 4-bit MLX copy of Cloudflare's clef-flash that turns a state and typed questions into a probability for every allowed answer on Apple Silicon Macs. Not a chat model. Runs locally with its own loader.

Made by mlx-community 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 is Cloudflare's clef-flash converted to 4-bit MLX for Apple Silicon. You give it a state, which can be text, JSON, images or video, plus a set of typed questions. It returns a probability for every allowed option in one pass.

It is not a chat model. Chat tools may load the backbone but produce meaningless text, so you use the loader file that ships with it, clef_mlx.py.

Who it suits: Developers with an Apple Silicon Mac who want to classify or route messages, images or other inputs locally, with structured answers instead of free text.

Industry | Who uses it | How they use it | Example
Customer support | Support engineers | Route incoming messages by team and urgency | Sending a checkout failure report to the technical team
Operations | Ops teams | Check whether a service is down from a report | Marking an outage message as urgent
Finance | Back-office developers | Check images such as receipts | Asking whether a receipt total is legible
Software | App developers | Add on-device decisions to a Mac app | A local triage step before a message reaches a person

Anyone looking for a chat model can skip it, and so can anyone without Apple Silicon.

Where to use it:

• Macs with Apple Silicon, using MLX.
• A Python script that imports the bundled loader.
• The command line, through the loader's predict command.
• A local server that mimics the SystemOne API.

Use case | Who | What they get back
Message triage | Support teams | A probability for each team or urgency option
Yes or no checks | Operations | A probability for each answer to a typed question
Image checks | Developers | Answers about an attached image

How to set it up: You need a Mac with Apple Silicon and enough RAM. The tested set is mlx at 0.32.3, mlx-lm at 0.32.0 and mlx-vlm at 0.7.4. The loader warns if you use an untested mlx-vlm minor version.

• A Mac with Apple Silicon
• Python
• 16 GB of Mac RAM at minimum for this 9B 4-bit variant
• No torch is needed

1. Install mlx-vlm and the Hugging Face library.
2. Download the model files with the snapshot call shown below.
3. Import clef_mlx from the downloaded folder and load the model.

pip install "mlx-vlm>=0.7.4,<0.8" huggingface_hub   # no torch needed

How to use it: Load the model, then call systemone with a state and a set of typed questions. The question types in the card are choice, score and noul.

import sys
from huggingface_hub import snapshot_download

path = snapshot_download("mlx-community/clef-flash-4bit")
sys.path.insert(0, path)
import clef_mlx

model = clef_mlx.load(path)
response = model.systemone({
    "model": "clef-flash",
    "state": "Our checkout started returning errors and orders are blocked.",
    "questions": {
        "department": {
            "type": "choice",
            "instructions": "Which team should handle the message?",
            "criteria": {"billing": "Payments or invoices", "technical": "Bugs or outages"},
        },
        "urgency": {"type": "score", "criteria": ["Can wait", "This week", "Today"]},
        "outage": {"type": "noul", "instructions": "Is a service down?"},
    },
})
print(response["answers"])

For images, pass PIL images in the images field. Videos go in as frame arrays in the videos field.

from PIL import Image
model.predict({
    "state": {"task": "Review the attached receipt."},
    "images": [Image.open("receipt.jpg")],
    "questions": {"legible": {"type": "noul", "instructions": "Is the receipt total legible?"}},
})

You can also run the loader straight from the command line, inside the downloaded folder:

cd "$(hf download mlx-community/clef-flash-4bit --quiet)"
python clef_mlx.py predict \
  --state "Checkout has been failing for every customer for the last hour." \
  --questions '{"urgent": {"type": "noul", "instructions": "Is this urgent?"},
               "team": {"type": "choice", "criteria": {"billing": "Payments", "technical": "Outages and errors"}}}'

To run a local server instead, use the serve command:

python clef_mlx.py serve --port 8000

Tool | What it does
clef_mlx.load | Loads the model and its question head from a folder
model.systemone | Answers typed questions about a state
model.predict | Takes a state, images and questions in one request
clef_mlx.py predict | Runs a prediction from the command line
clef_mlx.py serve | Starts a local server with a systemone endpoint

Example: Give the model a short report saying checkout is failing and orders are blocked. Ask which team should handle it, how urgent it is and whether a service is down. The answers field holds a probability for every option of each question.

For newcomers and experts: If you are new, think of it as a triage helper that answers multiple-choice questions about an input and shows how sure it is. Run the Python example once with the sample report.

For experts, it is a 9B base with a joint schema head, run in a single forward pass. The card lists 6.2 GB of download, about 7.0 to 8.6 GB peak memory from 1k to 16k tokens, and latency of 0.31 s at 1k tokens. Larger 27B variants score higher, while the 9B ones run much faster.

Provide a state and typed questions
The loader runs the 4-bit backbone
The joint schema head scores each option
One forward pass gives all probabilities
You read the answers field

Title: Download size in GB
clef-flash-4bit: 6.2 GB
clef-flash-8bit: 10.7 GB
clef-4bit: 16.3 GB
clef-8bit: 29.8 GB

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

• Memory: under macOS the GPU gets about three quarters of the RAM by default, so plan for more RAM than the peak figure.
• License: Apache 2.0.
• Cost to run: no API key or paid service is named.
• Data it reaches: your states and images stay on your Mac. The local server has no authentication and listens on the local machine only, so keep it that way unless you add your own proxy.
• Not for: chat, or anything that loads it as an ordinary text model. Requests run one at a time, and videos are not supported over the server.

On Hugging Face, 13 November 2026

• Task: zero shot classification
• Library: mlx
• Parameters: 9.4B
• Weights: about 37.6 GB in U32
• Formats: Safetensors
• Downloads in the last 30 days: 3,524
• Likes: 40
• License: apache-2.0
• Based on: Cloudflare/clef-flash
• Tags: safetensors, qwen3_5, clef, cloudflare, systemone, structured-output, classification, multimodal, custom-code, decision-model, zero-shot-classification, 4-bit

Links

• Model: https://huggingface.co/mlx-community/clef-flash-4bit
• Files: https://huggingface.co/mlx-community/clef-flash-4bit/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download mlx-community/clef-flash-4bit

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