Unlimited-OCR — parse a document locally

baidu/Unlimited-OCR (MIT, arXiv 2606.23050) is a 3B-total / 500M-activated MoE document parser. It reads a page image and returns markdown with LaTeX mathematics, tables, and a <|det|> bounding box for every block it found.

Self-Evolving Skill: improves through use. Every number below was MEASURED on this hardware, not copied from a README — provenance in ../../references/EMPIRICAL.md. If a measurement drifts, a flag stops working, or a new failure mode appears, fix this file and the references immediately rather than deferring. Re-verify before changing a documented fact.

The one rule: Free OCR. is the only prompt mode that works

Of the three prompt modes the upstream project documents, two are broken on MLX and both are refused by this CLI with exit 2 and a printed reason: