clef-flash-MLX-MXFP4
MLX conversion of Cloudflare/clef-flash for mlx-vlm. Clef is a decision model: it returns a probability for every option of every question in one forward pass and does not generate text.
Clef support is not yet in an mlx-vlm release:
pip install "git+https://github.com/Lazarus-931/mlx-vlm.git@feat/clef"
from mlx_vlm import load, predict
model, processor = load("nativ-community/clef-flash-MLX-MXFP4")
result = predict(model, processor, "Please refund my duplicate charge", {
"department": {
"type": "choice",
"instructions": "Which team should handle this ticket?",
"criteria": ["billing", "technical", "sales"],
},
})
print(result["answers"]["department"]["value"])
| Field | Value |
|---|---|
| Source | Cloudflare/clef-flash |
| Source revision | 17f0b0ad64efb65d273590632833508766b2aae6 |
| Quantization | mxfp4, group size 32 |
| mlx-vlm | Lazarus-931/mlx-vlm@c16f81aa |
| Verification | Identical token ids on 11/11 reference records (text, image, video, image+video, two images, max_pixels, fps, num_frames); same answer as Cloudflare's torch reference (fp32) on 25/25 questions; largest probability gap 0.0822 |