Saturate Library OCRs 14,483 Pages on One GPU for $1.47 via Hugging Face Jobs

Key Info

A developer re-OCR'd 14,483 pages of fossil fuel industry documents using the Saturate library on a single GPU for $1.47, running through Hugging Face Jobs at 8.9 pages/s.

Highlights

• Saturate runs models over a dataset or bucket against any OpenAI-compatible endpoint or inference engine.
• Concurrency was not manually tuned; Saturate settled on 512 by itself.
• The run achieved 8.9 pages/s once running via Hugging Face Jobs.
• A sample comparison shows existing OCR output versus Xiaomi-OCR-0 on one page line.