Coming soon
Unflatten W
PDF / document images → editable DOCX
LabsResearch preview
Unflatten W
From PDF and document images to editable DOCX
Unflatten W is a research-driven document intelligence model focused on reconstructing structured Word documents with high fidelity. It is designed to understand visible document layout, structure and content, then rebuild them as clean, editable DOCX files.
Research drivenReal-world documentsStructured output
Unflatten W
Coming soonPDF / document images
Canonical AST
Structured document representation
Editable DOCX
PDF / document images
Scanned documents, reports, contracts and forms
Editable DOCX
Structured, editable and renderer-ready
- Layout analysis
- Structure reconstruction
- Content alignment
- Typed properties
Introducing the Unflatten family
Unflatten is a research family of document intelligence models focused on turning complex files into editable, structured outputs. Each model targets a different office-document workflow while sharing a common foundation in layout understanding, structured decoding and reconstruction quality.
Coming soon
Unflatten X
Documents and tables → structured spreadsheet output
Coming soon
Unflatten P
Slides and visual layouts → editable presentation structure
Coming soon
Shared foundation
Layout understanding, structured decoding and model research
What Unflatten W is designed to do
Coming soonLayout reconstruction
Paragraphs, headings, lists, sections and multi-column document structure.
Table recovery
Dense tables, merged cells, spans and structured table reconstruction.
Image & resource handling
Inline images, document assets and image-resource recovery.
Dense & MoE architecture
Future research will explore dense and Mixture-of-Experts (MoE) variants built on a shared document core.
Coming soon
The Unflatten model family
A single research direction for editable Office reconstruction: Word, spreadsheets and presentations — with a shared document-understanding foundation.
Coming soon
Technical summary
The current Unflatten W architecture separates learned document reconstruction from deterministic assembly and rendering. The model predicts document structure, visible text and typed properties; deterministic components promote validated output to the Canonical AST and render it to DOCX.
- Canonical AST
- Qwen3-VL backbone
- Structured decoding
- DOCX renderer
- Research preview
Pixels
PDF / image input
Model
Qwen3-VL + document reconstruction branch
Validated representation
Canonical AST → deterministic DOCX renderer
Research status
Coming soonUnflatten W and the wider Unflatten family are under active development at Evoliq Labs. This page is a public research preview, not a production-availability announcement.