IBM Research's SmolDocling, a 256M-parameter vision-language model, delivers fast document OCR and multimodal processing at 0.35s per page on consumer GPUs, handling text, formulas, code and charts efficiently.