INT8 quantized vision models for real-time robotic perception. SAM2, DINOv2, CLIP, SigLIP, Depth Anything.