MonkeyOCRv2 pushes doc-parsing SOTA with 11× smaller encoder

A paper posted this week is worth pausing on if you work anywhere near document AI, because it argues something the field has been quietly hoping was true: for parsing documents, a small model trained on a huge amount of the right data can beat a much bigger general-purpose one.

The system, MonkeyOCRv2, is a 0.7B parameter document-parsing model whose vision encoder is roughly 11× smaller than the 3B baseline it is being compared against. The authors report a 2.8% absolute improvement on MDPBench and describe wins over CLIP, DINO, and SAM variants across eight document understanding benchmarks. The training recipe is two-part: image-to-text generation paired with pixel-level document reconstruction, aimed at holding on to character detail and layout at the same time.

The bigger claim sits underneath the model itself. The paper introduces MonkeyDoc v2, a pretraining corpus of 113 million images spanning 17 languages, and evaluates across five document tasks: text recognition, formula recognition, text detection, document tampering detection, and overlapping text segmentation. That corpus scale is the real bet, that a domain-specific pretraining set at general-VLM scale moves you further than a bigger, more general encoder.

The honest caveats: this is a fresh preprint that has not been externally reproduced yet, the dataset licensing and openness are not spelled out on the abstract page, and benchmark wins on curated document sets do not always survive the noisy PDFs and handwritten scans that show up in real enterprise pipelines. What the paper does not tell you is how throughput and latency compare against the OCR stacks banks and insurers actually run, or which of the 17 languages are well-covered versus token inclusion.

If it holds up in independent testing, the direction is where the interesting money is. Teams doing high-volume OCR on invoices, contracts, and medical records get a real shot at cutting inference cost without giving up accuracy, and the same pretraining recipe may travel to other structured-image domains like forms and technical diagrams.