Skip to content
Other

Unflatten W Dense 1.0

Turn Page Scans and Photos into Editable Word Docs

3 min read

Developed by EVOLIQ s.r.o., this model goes beyond raw text extraction to reconstruct a page's logical layout from a raster image—capturing paragraphs, text blocks, tables, and image placements—before a custom deterministic renderer builds a ready-to-edit .docx file. Under the hood, it pairs a frozen Qwen3-VL-4B-Instruct VLM (vision-language model) backbone with trained reconstruction modules and typed property heads, released here as a checkpoint trained with 48 steps of reinforcement learning (RL). Unlike standard OCR that outputs flat text, the system relies on grammar-constrained decoding to assemble a full document hierarchy with native Word properties. It's built for local document and PDF digitization without sending sensitive files to cloud services.

Key features

  • Qwen3-VL-4B-Instruct architecture: Extracts visual features from pooler_output and DeepStack layers (5, 11, and 17), routing them into specialized document reconstruction modules.
  • Grammar-constrained decoding: Structure generation is guided by a composite document grammar, while the deterministic Unflatten renderer compiles the final .docx file.
  • Ready-to-run local bundle: Ships alongside the weights with the Unflatten W app (featuring a web UI, headless API, and Swagger UI) plus the runtime source archive unflatten-runtime.tar.gz.
  • Offline execution: Includes the exact frozen backbone, tokenizer, and image processor in the repository for standalone loading with trust_remote_code=True.

What you can do

  • Convert PDFs and scans to Word locally: Deploy the Unflatten W app on a CUDA-enabled NVIDIA GPU to turn document pages into editable .docx files.
  • Integrate into custom processing pipelines: Connect to the local server via a Bearer-authenticated REST API (documented via Swagger UI on port 8421) for batch-processing multi-page PDFs.

Example

The single-page processing pipeline follows this flow:
Page image → Unflatten W Dense → structured document → Unflatten renderer → editable Word
Once the local Unflatten W app is up and running, you can upload a scan or PDF through the browser UI or send API requests via Swagger UI at [http://127.0.0.1:8421/docs](http://127.0.0.1:8421/docs) (evaluation was performed on 300 DPI page images).

Read moreDetails
100-page validation: 90/100 passed the text tolerance check, 99/100 produced a valid diagnostic Word file, with a CER of 13.8949% and a WER of 8.1013%.
Hardware & requirements: Requires an NVIDIA GPU with CUDA (tested on DGX Spark) and the custom Unflatten runtime; standard Transformers text-generation pipelines are not supported.
Limitations: Single-page research release (product_quality_claim: false); errors in text, reading order, and layout may occur.
License: Free for personal non-commercial use only (EVOLIQ Personal Use and Commercial License 1.0); commercial deployment requires a paid license, while the base Qwen model is under Apache 2.0.
Discuss Telegram 0 2

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 8.3 GB · ready files9.6 GB
  • FP8 / INT8recommendedweights 4.1 GB5.0 GB
  • 4-bitweights 2.3 GB3.1 GB

Tip: go 8-bit.FP8 / INT8 is within about a percent of full precision in quality, takes half the memory and runs faster on RTX 40/50 and H100. Full precision only pays off for fine-tuning.

What to get for this model

Massed ComputeOpen Massed Compute

Your own GPU: FP8 / INT8 — from RTX 2060 · BF16 — from RTX 3080

Massed Compute hourly prices, November 2026. The discounted price applies with the coupon at checkout.

Estimate: weights plus working memory. Long videos, high resolutions and long contexts need more.

Run it on Massed Compute

The script downloads the model to the rented machine. How to run it is in the model card.

  1. 1. Rent a machine with a fitting GPU

    Open Massed Compute
  2. 2. Paste the script into the machine's terminal

    Connect over SSH (the Massed Compute panel shows the command) or open a terminal on the machine's desktop.

    HF_TOKEN / CIVITAI_TOKENThe script starts with token lines: paste yours between the quotes. The Hugging Face token is for gated models, the Civitai one for Civitai models.

    # Evoliq/Unflatten-W-Dense-1.0
    mkdir -p ~/models
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv tool install "huggingface_hub[cli]"
    hf download Evoliq/Unflatten-W-Dense-1.0 --local-dir ~/models/Unflatten-W-Dense-1.0
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to Unflatten-W-Dense-1.0
  3. 3. Use it

    The model is in ~/models. How to run it is in the model card (link above).

Stop the machine in the panel when you are done: it bills by the minute while it runs.

How to speed it up

Picked for this model type. Choose your hardware above to see what fits you exactly.

  • FlashAttention 44.0.0b40from November 18, 2026new release

    fastest exact attention on B200

    Kernels for Blackwell (beta). Exact attention without quantisation — fine for everything, LLMs included.

    How to enable: pip install flash-attn-4; vLLM and SGLang pick it up themselves.

    Earlier generations (2)
    • FlashAttention 3 — The Hopper (H100/H200) version with FP8.
    • FlashAttention 2 — The standard for RTX 30/40/50 and A100: exact attention without the memory cost of long contexts.
  • torch.compile2.15.0from October 28, 2026new release

    10–30% for free

    Compiles the model for your GPU. Slower first run, faster afterwards.

    How to enable: model = torch.compile(model); in ComfyUI the TorchCompileModel node.

Model family

New posts about versions of Qwen3-VL-4B-Instruct (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 12All on Hugging Face
Fine-tunes 12All on Hugging Face
LoRAs and adapters 5All on Hugging Face

About the model

Evoliq/Unflatten-W-Dense-1.0
Author
Evoliq
Category
Other
Released
Posted

On NeyroLearn

Hugging Face

20

Ask AI about this model

Answered by Gemini 4 · the chat is kept only in this tab

Comments