rakedoc-nano-2

rakedoc-nano-2

Parses document page images into text, tables and charts with markdown formatting. A 1.2B-parameter model with better chart extraction, for developers building document pipelines.

Made by cloudraker on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: rakedoc-nano-2 is a 1.2B-parameter document parser on the Qwen2-VL architecture. CloudRaker fine-tuned it from KDL-Frontier-Parser-nano-v2 to improve chart extraction and text formatting. The tuned weights are already merged into the model in bf16, so there is no adapter to load separately. It powers the document parsing pipeline of CloudRaker's Paperwork API.

Who it suits: It suits developers and data teams who parse page images into structured text and want a small model they can self-host.

Title: Who might use it
Industry | Who uses it | How they use it | Example
Back-office operations | Data teams | Turn page images into structured markdown | A report page that holds a table and a chart
Document AI | Developers | Self-host a small parser behind their own service | Replacing KDL-Frontier-Parser-nano-v2 in an existing setup
Benchmarking | Researchers | Score document parsers on ParseBench | Reading the overall, table and chart scores

People who need highlighted text or who work outside report-style documents can skip it, since the card says highlights are never emitted and other domains are untested.

Where to use it:

• A vLLM server: the card gives a serve command
• A single GPU: the card's benchmark runs used one
• ParseBench: the card shows how to run its pipeline against a local endpoint
• Windows, macOS or Linux: download the files with Python

Title: Jobs it does
Use case | Who | What they get back
Page parsing | Developers | Text, tables and charts read from a page image
Table extraction | Data teams | Table output in the OTSL format
Chart extraction | Analysts | Chart content from report pages

How to set it up: You need these first:

• Python on Windows, macOS or Linux
• vLLM, to serve the model
• A GPU with room for about 2.3 GB of BF16 weights

1. Install the Hugging Face hub package.
2. Download the model files.
3. Start the model with vLLM using the card's serve command.

pip install -U huggingface_hub
hf download cloudraker/rakedoc-nano-2

The card gives this command to serve the model. You can swap it in wherever KDL-Frontier-Parser-nano-v2 was running.

vllm serve cloudraker/rakedoc-nano-2 --served-model-name rakedoc-nano-2 \
  --max-model-len 8192 --gpu-memory-utilization 0.85 --max-num-seqs 24 \
  --trust-remote-code --limit-mm-per-prompt '{"image":1}'

How to use it: Send page images to the running server with the stage prompts that KDL-Frontier-Parser-nano-v2 uses. The model reads one image per request.

Title: Stage prompts
Tool | What it does
Layout Detection | Finds the layout of a page
Text Recognition | Reads the text, including formatting marks
Table Recognition | Reads tables and answers in OTSL output

Example: You serve the model with vLLM and send one page image with the Table Recognition prompt. A table comes back in OTSL output.

For newcomers and experts: If you are new to this, think of it as a reader that looks at a picture of a page and writes out its text, tables and charts. Start by serving it and sending one clean page.

If you work in the field, The model averages the weights of two separate tunes of KDL v2, at 0.7 and 0.3. One tune targets formatting marks through GRPO on text recognition. The other targets charts with a short SFT warm start and then GRPO on the chart stage. The card says ParseBench was used only to pick between finished candidates.

Send a page image with a stage prompt
The model answers layout, text or table
The pipeline writes markdown output

Good to know: The card reports the mean of 3 runs on ParseBench and compares it with KDL-Frontier-Parser-nano-v2.

Title: ParseBench overall score
rakedoc-nano-2: 80.09 points
KDL v2 with the same pipeline: 79.74 points
KDL v2 with its own pipeline: 79.06 points

• License: agpl-3.0, inherited from the KDL weights, with LICENSE and NOTICE files in the repository
• Cost to run: none named, though the card mentions CloudRaker's Paperwork API, which this model powers
• Data it reaches: the page images you send to the server you run
• Not for: highlighted text, since it is never emitted, or domains beyond ParseBench and synthetic data, which are untested

On Hugging Face, 13 November 2026

• Task: image text to text
• Library: transformers
• Parameters: 1.2B
• Weights: about 2.3 GB in BF16; about 578 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 543
• Likes: 6
• License: agpl-3.0
• Based on: KDLAI/KDL-Frontier-Parser-nano-v2
• Quantized versions: 1
• Tags: transformers, safetensors, qwen2_vl, image-text-to-text, document-parsing, table-extraction, chart-extraction, parsebench, conversational

Links

• Model: https://huggingface.co/cloudraker/rakedoc-nano-2
• Files: https://huggingface.co/cloudraker/rakedoc-nano-2/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download cloudraker/rakedoc-nano-2

Then follow the usage in its model card. Nothing in this kit runs by itself.