Optimum Intel adds OpenVINO Runtime support and NNFC quantization for Transformer models

AI Impact Summary

Intel OpenVINO is now integrated into Optimum Intel, enabling OVRuntime-based inference for Transformer models on Intel hardware with the OVQuantizer and NNNCF tooling. The release exports OpenVINO models as XML/BIN and supports post-training static quantization for encoder models, with initial ViT benchmarks showing up to 3.8x memory reduction and ~2.4x lower latency. Quantization requires a calibration dataset and a warmup/compilation step, and Encoder-Decoder quantization remains disabled until the next OpenVINO release. Hosting can be done from the Hugging Face hub or locally, broadening deployment options for edge and data center environments.