RF-DETR base — MLX (fp32)

Roboflow's RF-DETR base detector converted to MLX safetensors for Apple Silicon. COCO classes (91), float32.

Licensing follows Roboflow's RF-DETR release; see the upstream project for terms.

Usage

from mlx_vlm import load
from mlx_vlm.extraction import extract

model, processor = load("nativ-community/rfdetr-base-fp32")
result = extract(model, processor, "photo.jpg")
print(result["boxes"], result["scores"], result["labels"])

From the command line:

mlx_vlm.extract --model nativ-community/rfdetr-base-fp32 \
    --image photo.jpg --output boxes.npz

Conversion

python -m mlx_vlm.models.rfdetr.convert --variant base --output rfdetr-base-fp32

The converter downloads the official weights from Roboflow and needs only torch plus safetensors; inference needs neither.

Verification

On COCO val2017/000000039769.jpg this checkpoint returns two cats (0.93, 0.89), two remotes, and a couch — matching the image contents.

Downloads last month: 25

Safetensors

Model size

32.2M params

Tensor type

F32

MLX

Hardware compatibility

Quantized