Core Epoch

INT8 ONNX computer vision models, quantized with Kenosis by Core Epoch. Each is measured against its FP32 original: 58% to 68% smaller, up to 70% faster on one CPU thread (measured on a CPU with AVX-VNNI), and keeping 94.9% to 99.96% of the original accuracy. Every file runs on both ONNX Runtime and OpenVINO.

Custom model quantization and licensing: coreepoch.dev/kenosis · licensing@coreepoch.dev