Accuracy-Energy Trade-offs for On-Premise Information Extraction
September 28, 2026
Small models (≤8B) show varying efficiency when processing page images versus parsed text depending on document layout. Results indicate batching reduces energy per page by 38-85% with no accuracy loss, while neural OCR consumes 17x more energy than classical OCR.
HOW THIS AFFECTS YOU
●
builderYou can optimize local deployment costs by prioritizing batching and classical OCR over neural alternatives for layout-rich documents.