INT8 Quantization Service

Kenosis

Kenosis is Core Epoch's quantizer for ONNX computer vision models. It converts a model to 8-bit integers (INT8) without retraining, and the same file runs on ONNX Runtime and OpenVINO. The models below are free to download, and we quantize customers' models as a service.

01 // PUBLISHED MODELS

Measured Against the Original Models

Every model below was quantized with Kenosis by Core Epoch, with no retraining, and keeps 94.9% to 99.96% of its original accuracy. Speeds are ONNX Runtime on one thread, measured on a CPU with AVX-VNNI.

Object Detection

RF-DETR Base

Detection Transformer · COCO val2017

Baseline: 53.32 AP (FP32)
Kenosis INT8 53.12 AP (99.6% retained)
Size: 115.1 MB → 41.9 MB (64% smaller)
Speed: 67% faster on 1 CPU thread
Object Detection

RT-DETRv2-S

Detection Transformer · COCO val2017

Baseline: 48.13 AP (FP32)
Kenosis INT8 45.67 AP (94.9% retained)
Size: 81.0 MB → 32.7 MB (60% smaller)
Speed: 70% faster on 1 CPU thread
Image Classification

EdgeNeXt-S

Hybrid ConvNet · ImageNet-1K

Baseline: 81.565% top-1 (FP32)
Kenosis INT8 81.535% (99.96% retained)
Size: 22.5 MB → 7.16 MB (68% smaller)
Speed: 32% faster on 1 CPU thread
Image Classification

XCiT-Tiny-12-p8

Vision Transformer · ImageNet-1K

Baseline: 81.22% top-1 (FP32)
Kenosis INT8 81.16% (99.93% retained)
Size: 27.0 MB → 8.59 MB (68% smaller)
Speed: 36% faster on 1 CPU thread
60% to 68% Smaller files than the FP32 originals, for the four models above
70% Faster RT-DETRv2-S on one CPU thread
0.06 Top-1 points lost, at most, by the two ImageNet models above
5 Models free to download on Hugging Face
See the Models on Hugging Face →

Quantize Your Model

Send us an ONNX vision model and we return an INT8 version. The fee is $1,500 per model, and you pay only if the result meets the accuracy target you set.

Quantization service: quantize@coreepoch.dev
License Kenosis for your own team: licensing@coreepoch.dev
U.S. patent applications pending