Run Vision-Language Models locally on Intel CPUs with OpenVINO and SmolVLM2-256M-Video-Instruct

AI Impact Summary

Intel-enabled workflow enables running Vision-Language Models locally on CPUs using Optimum-Intel and OpenVINO, demonstrated with SmolVLM2-256M-Video-Instruct. This reduces cloud dependency and data exposure, with potential latency benefits on devices without GPUs. The guide details two quantization paths (weight-only and static) and a complete export-to-IR process, but quantization can impact accuracy and requires validation before production deployments.