CoreEpoch/rfdetr-base-int8-onnx

RF-DETR Base: INT8 ONNX, 53.1 AP COCO in 41.9 MB

RF-DETR Base (Roboflow, Apache-2.0) quantized with Kenosis: 128-image COCO calibration, no retraining, weights and activations both INT8. 53.1 AP (99.6% of FP32) from a single 41.9 MB file, measured on CPU with stock ONNX Runtime and OpenVINO. No accelerator required.

Accuracy

COCO val2017, 4,800 images (val2017 minus the 200-image pool the 128 calibration images came from; list in eval_images.json), pycocotools bbox AP50:95.

model	ONNX Runtime	OpenVINO	size
FP32 baseline	53.32	53.32	115.1 MB
Kenosis INT8	53.12	53.05	41.9 MB

Speed

Against the FP32 model, measured on a CPU with AVX-VNNI: 67% faster on ONNX Runtime with one thread, 45% faster with four threads, and 91% faster on OpenVINO with four threads.

Run

from huggingface_hub import hf_hub_download
import numpy as np, onnxruntime as ort
from PIL import Image
path = hf_hub_download("CoreEpoch/rfdetr-base-int8-onnx", "rfdetr_base_560_int8_kenosis.onnx")
sess = ort.InferenceSession(path, providers=["CPUExecutionProvider"])
img = Image.open("your_image.jpg").convert("RGB")
x = np.asarray(img.resize((560, 560), Image.Resampling.BILINEAR), np.float32) / 255.0
x = ((x - [0.485, 0.456, 0.406]) / [0.229, 0.224, 0.225]).transpose(2, 0, 1)[None]
dets, labels = sess.run(["dets", "labels"], {"input": x.astype(np.float32)})
scores = 1.0 / (1.0 + np.exp(-labels[0]))
q, c = divmod(int(scores.argmax()), scores.shape[1])
print(f"COCO category {c}, score {scores.max():.2f}, box {dets[0][q]}")

Input 1x3x560x560, RGB, /255 then ImageNet mean/std. Outputs dets [1,300,4] (normalized cxcywh) and labels [1,300,91] (logits indexed by COCO category id). run_detect.py prints the detections for one image; eval_coco.py reproduces the accuracy table.

Integrity & license

rfdetr_base_560_int8_kenosis.onnx (41,946,555 B), SHA-256 40B0680FA56308CE6C0F8BF779CD069876D6F2A303B00B2C5FE5D1A29B19165E. Apache-2.0 (base RF-DETR Base by Roboflow, Robinson et al., "RF-DETR: Neural Architecture Search for Real-Time Detection Transformers", ICLR 2026, arXiv:2511.09554, retained). FP32 source exported with rfdetr 1.8.3 at 560x560, ONNX opset 17. The license covers these model files only; it grants no rights in Kenosis, Core Epoch's quantizer, which remains proprietary.

For licensing inquiries or custom model quantization: coreepoch.dev/kenosis · licensing@coreepoch.dev