Unlimited OCR vs DeepSeek OCR: document parsing review

Unlimited OCR is presented as a document parsing model for long documents and RAG pipelines. The speaker describes 3 billion total parameters, with 500 million active during inference, and a 32,000-token context. They report a 93.23% score on OmniDoc Bench V1.5, six percentage points above DeepSeek OCR. These are reported results, rather than independent validation.

The explanation centers on keeping the source document in context while retaining a sliding window of generated text. The speaker argues that this approach reduces the need to stitch page outputs together before indexing text for retrieval. Their comparison places DeepSeek OCR's emphasis on compressing document images into fewer visual tokens, while Unlimited OCR addresses cache growth during output generation.

The demo covers mathematical formulas, tables, reading order and handwriting. The speaker reports good results for formulas and tables, and cites the technical report for handling complex layouts. Handwriting exposes a clear limitation: mixed notes lose content, contain incorrect text and need substantial manual correction.

For developers evaluating open source document extraction, the speaker cites an MIT license and downloadable weights. The video provides an architectural overview and chatbot examples, but does not establish hardware requirements or show a local installation.