ONNX INT8 Quantization Engine

Kenosis

Kenosis quantizes ONNX vision models to portable INT8 — the same artifact runs on stock ONNX Runtime and OpenVINO. The output is standard ONNX with no custom ops or proprietary formats. INT8 quantization is what makes a model small and fast; Kenosis keeps the accuracy while it does. A separate set of Kenosis-quantized models is published on Hugging Face, where anyone can download them and check the results themselves. Four U.S. patent applications are pending.

The same model, quantized two ways

Kenosis
81.5%
ONNX Runtime static quantizer
33.7%

EdgeNeXt-S, ImageNet-1K top-1 on the full 49,872-image validation set — the same model and the same calibration images, quantized by ONNX Runtime's own static quantizer at its default calibration. See the full numbers →

Try the models free →
−0.03 pts
Top-1 classification change from FP32 on that same full validation set. The confidence interval includes zero, so the quantized model is statistically indistinguishable from the original.
7.16 MB
One standard ONNX file holding that 81.5%, down from 22.5 MB at full precision. It runs on OpenVINO from the same bytes, and across a 2,000-image comparison the two runtimes agree within the margin of error
93–94%
COCO detection mAP retention on PP-YOLOE+ across 320/416/640, on ONNX Runtime
48.4 vs 28.9
RF-DETR mAP — Kenosis vs ONNX Runtime's static quantizer at its default calibration

Licensing Inquiries

Kenosis is proprietary software available under commercial license from Core Epoch LLC. For licensing terms, evaluation access, and partnership opportunities:

core@coreepoch.dev
Research publication: Fusion-Aware QDQ Placement