Handwritten Arabic archives, made searchable
Manuscripts and historical documents, made searchable
Archives holding handwritten manuscripts and degraded historical documents need recognition technology trained for exactly that material - highly cursive handwriting and archival Arabic - not general-purpose OCR.
Why general OCR stops at the first manuscript
Highly cursive handwriting on degraded archival paper is a different problem from printed text, and general-purpose recognition does not degrade gracefully on it. It fails. Baseer Nakba is an open-source vision-language model built specifically for Arabic handwritten text recognition on historical manuscripts, and it took first place at the NAKBA NLP 2026 Arabic Manuscript Understanding shared task.
Built for the material, not adapted to it
Baseer Nakba is trained on highly cursive handwritten Arabic from degraded archival documents, the exact case where general recognition models stop being useful.
Layout is part of the record
Baseer preserves columns, tables and hierarchy in Markdown rather than flattening a page into a line of characters, so a transcription keeps the structure the original had.
Open weights, published work
The manuscript model is released openly and documented in a paper, so an archive can evaluate it before committing a collection to it.
What archives deploy
- Manuscript recognition
Cursive handwriting on degraded paper, transcribed rather than approximated.
Baseer Nakba is a vision-language model for Arabic handwritten text recognition, optimised for historical manuscripts. Its parent model, Baseer, handles printed and mixed archival material and returns structured Markdown that preserves the page's own layout, so the collection comes out as documents rather than as a pile of text files.
Document Intelligence - Making the collection searchable
A transcribed archive is only useful once it can be queried.
Structured Markdown is the input a retrieval system can actually index. With Arabic retrieval handled by the Kawn-Embed family, a catalogue becomes searchable by meaning rather than by exact string, and an answer points back at the document it came from.
Enterprise Search & RAG
Deployment options
The manuscript model is released with open weights, and the wider document stack deploys in the cloud, on-premises, or on dedicated servers inside the Kingdom, so a collection that cannot leave the building does not have to. The Trust Center sets out the deployment options in full.
Let's talk about what you're trying to build.
Tell us the problem. We'll tell you honestly whether AI is the right answer.