mlx-community/clef-4bit

Cloudflare/clef converted to MLX (4-bit) for Apple Silicon.

Clef turns a state (text, JSON, images, or video) plus a schema of typed questions into a probability for every allowed option, in a single forward pass. It is not a chat model — mlx_vlm.generate, mlx_lm.generate, and LM Studio will load the backbone but produce meaningless text. Use the bundled clef_mlx.py loader, which runs the backbone and the joint schema head.

Which variant fits your Mac?

Variant	Base	Download	Peak memory (1k / 4k / 16k tokens)	Latency (1k / 16k tokens)	Minimum Mac RAM
clef-flash-4bit	9B	6.2 GB	7.0 / 7.2 / 8.6 GB	0.31 s / 7.0 s	16 GB
clef-flash-8bit	9B	10.7 GB	11.4 / 11.6 / 13.0 GB	0.34 s / 7.7 s	24 GB (16 GB for short prompts)
clef-4bit (this repo)	27B	16.3 GB	17.1 / 17.5 / 19.6 GB	1.4 s / 26.0 s	32 GB
clef-8bit	27B	29.8 GB	30.5 / 30.8 / 33.0 GB	1.5 s / 32.0 s	48 GB

Measured on an M5 Max with text input. macOS lets the GPU use only about 70–75% of RAM by default, so the minimum RAM is higher than the peak. The 9B variants are much faster; the 27B variants score higher (see the quality check below).

Usage

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

Tested with mlx 0.32.3, mlx-lm 0.32.0 and mlx-vlm 0.7.4. clef_mlx.py uses mlx-vlm internals, so it warns if you load it with an untested mlx-vlm minor version.

import sys
from huggingface_hub import snapshot_download

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

model = clef_mlx.load(path)
response = model.systemone({
    "model": "clef",
    "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"])

Images (PIL) and videos (frame arrays) go in images / videos, as in the original:

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?"}},
})

See the original model card for the input format, question types, and benchmarks.

Try it from the command line

clef_mlx.py runs on its own. From the downloaded repo it uses that repo by default; elsewhere pass --model mlx-community/clef-4bit.

cd "$(hf download mlx-community/clef-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"}}}'

--questions takes JSON, a .json file, or - for stdin; --request takes a whole request body instead; --image photo.jpg attaches an image (repeatable). The output is the SystemOne response as JSON.

Run a local SystemOne server

python clef_mlx.py serve --port 8000          # http://127.0.0.1:8000, local only by default
curl http://127.0.0.1:8000/v1/systemone -H "Content-Type: application/json" -d '{
  "model": "clef",
  "state": "Checkout has been failing for every customer for the last hour.",
  "questions": {"urgent": {"type": "noul", "instructions": "Is this urgent?"}}
}'

• POST /v1/systemone takes the same request body as the Jev/SystemOne API and returns the same response (plus usage.latency_ms), so existing SystemOne clients can point at your laptop. GET /health and GET /v1/models are also available.
• Images go in images as data URLs, base64, or http(s) URLs. Videos aren't supported over HTTP.
• Errors return JSON: 400 for invalid requests, 413 with "maximum context length" for inputs that don't fit. Add "truncate": false to a request (or start with --no-truncate) to get the 413 instead of truncation.
• Requests run one at a time on the GPU. There is no authentication: keep the default 127.0.0.1 binding unless you put your own proxy in front.
• For the Decision Index, start with --no-truncate and use its http engine: python -m decision_index run --engine http --option base_url=http://127.0.0.1:8000.

Limitations

• Not a chat model. Text generation tools (mlx_lm.generate, mlx_vlm.generate, LM Studio, Ollama) load the backbone but give meaningless output. Only clef_mlx.py runs the decision head.
• Long inputs are truncated by default, exactly like the reference implementation: the state is cut so the whole prompt fits in 16,384 tokens. Pass truncate=False to predict()/systemone() to get a clef_mlx.ContextTooLong error instead, or raise max_length (the head was trained at 16k).
• One record per call. There is no batching; requests run one after another.
• Images or videos, not both in one record. Several images, or several videos, are fine.
• Speed depends heavily on the chip. Latencies above are from an M5 Max; long prompts on base M-series chips will be several times slower.
• Custom code. Like the original release, this repo ships Python (clef_mlx.py) that you import and run. Read it before use if that matters in your environment.

Conversion

• Backbone: mlx_vlm.convert -q --q-bits 4 --q-group-size 64 (vision tower kept in bf16).
• Joint schema head: joint_head.safetensors copied unchanged (bf16) and run by clef_mlx.py.
• processor_config.json is the original from Cloudflare/clef; prompt/token layout matches the reference joint_schema_model.py exactly (images and video).

Parity vs. official PyTorch implementation (bf16)

Inputs	Top answer agrees	Max abs Δprob
Text (4 records, 10 questions)	10/10	0.037
Images + video (5 records, 9 questions)	9/9	0.097

Measured on an M5 Max (128 GB). Small spot-check, not a full benchmark run.

Quality check: Decision Index (sampled)

	Decision Index (sample)	Same top answer as 8-bit	Mean max abs Δp	Median latency
This model (4-bit)	57.02	98.1% (15,915 answers)	0.023	1.03 s
MLX 8-bit, same rows	57.96	—	—	1.13 s
Cloudflare published (full suite)	61.21

4-bit costs about 1 index point vs 8-bit on identical rows (8-bit tracks the PyTorch reference within 0.007 probability on spot checks). Most of the gap to the published score is already present at 8-bit — sample noise, harness differences, and Arts & Human Taste in particular — not quantization. Latency is per request on Apple Silicon (median request ~1k tokens); MLX prefill is ~2k tok/s for the 9B, slower for the 27B.

Method: Decision Index 0.2.1 kit (suite rebuilt byte-identical), stratified 2,000-request sample across all 44 benchmarks (suite sample --n 2000), engine = clef_mlx.py with max_length=16384 and no truncation (over-length requests are refused and count as wrong; 15 of 2,000, mostly BRIGHT). The index is computed from each benchmark's native metric on the sampled rows with the kit's chance correction and weights; HLE and iSarcasmEval are set to 0 to match how Cloudflare's published run is scored. With ~40 rows per benchmark, per-benchmark numbers are noisy (±10+ pts) — only the index is meaningful.

License

Apache-2.0, following Cloudflare/clef.