Introducing rakedoc-nano: a state-of-the-art open-weight VLM for document parsing

Introducing rakedoc-nano: a state-of-the-art open-weight VLM for document parsing

A 1.2B-parameter vision parser behind the CloudRaker Paperwork API. State-of-the-art among open-weight parsers, bounding boxes on every element, and enterprise pricing of $0.60 per 1,000 pages.

Back to journal
Highlights
  • #1 among VLMs on ParseBench. rakedoc-nano scores 77.2 overall on the ParseBench leaderboard: first among all vision-language models, proprietary frontier models included, and fifth overall, at a fraction of the price of everything above it.
  • Spatially aware, all the way down. Bounding boxes for every element, text blocks and table cells included, not only images and tables. Structure survives into retrieval, grounding, redaction, and audit.
  • Open weights. Published on Hugging Face under AGPL-3.0, so you can run it, inspect it, fine-tune it, and keep documents inside your perimeter.
  • $0.60 per 1,000 pages at enterprise volume. Rack rate is $2 per 1,000 pages, charged only on the pages that need a model — three in four pages route through the free rule-based tier, all-in.
  • Runs anywhere. Available today in the Paperwork API and your CloudRaker Workspace, or self-hosted with vLLM on any GPU with 8 GB of memory.

Today, on day one of CloudRaker Launch Week, we are releasing rakedoc-nano, the compact document parsing model behind the Parse capability of the CloudRaker Paperwork API. It turns PDFs, scans, and filings into clean Markdown, structured tables, and per-element bounding boxes. Starting now it is yours too: open weights, a published benchmark harness, and a pipeline priced for running on everything.

We build the agentic paperwork stack for the enterprise: insurance, healthcare, legal, financial services. The documents that matter most to our customers are the ones with a 30-column rate table and a three-level header, and the ones where an auditor asks where a number came from. That is exactly where we focused, and it shows in the numbers. Used as a pre-processing step, rakedoc-nano also cuts the token count downstream models need to complete a workflow, and its bounding boxes let them locate any word on the page.

Performance

rakedoc-nano is the state-of-the-art open-weight document parser VLM on ParseBench. It scores 77.2 overall: first among all vision-language models, proprietary frontier models included, and fifth overall, behind only LlamaParse's three managed tiers and oi-parser. Every number in this post comes from the ParseBench leaderboard, including our submitted run (PR #120), which places rakedoc-nano in the most attractive quadrant of the price-performance map: frontier-adjacent accuracy at a commodity price.

CloudRaker (filled = enterprise; hollow = rack rate)Others
Figure 1. ParseBench overall score versus price per 1,000 pages, log scale, with the free tier at $0. The filled dot is the $0.60 enterprise price; the hollow dot is the $2 rack rate. Scores and prices from the ParseBench leaderboard; models without a published price are omitted (oi-parser, fourth overall at 78.3, reports no price).

rakedoc-nano is strongest where enterprise documents are hardest: Semantic Formatting (headings, bold, superscripts, the structure that makes a parsed document usable downstream), where it scores 71.7 — sixth on the whole leaderboard, third among VLMs, level with Opus 4.8 — and Content Faithfulness, at 88.8 within 3.5 points of the best score on the bench. On the three dimensions that matter most for filings (Tables, Content Faithfulness, Semantic Formatting), it is ahead of every dedicated parsing API except LlamaParse's agentic tiers.

Figure 2. ParseBench overall score (mean of five dimensions), from the leaderboard including our submission. The free tier is an estimate for the rule-based pipeline.

Per dimension

Model Overall Tables Charts Content Semantic Grounding
LlamaParse Cost Effective 80.6 84.2 77.9 89.9 67.3 83.8
rakedoc-nano 77.2 86.4 64.9 88.8 71.7 74.3
florin-parser-nano (open) 76.7 86.1 65.2 87.4 70.6 74.1
Gemini 3 Flash (Thinking High) 75.1 91.5 64.8 90.9 68.3 59.8
Reducto (Agentic) 73.0 80.4 73.4 86.4 57.6 67.1
Fable 5 70.8 89.8 52.2 90.0 72.6 49.2
Chandra OCR 2 (open) 70.1 89.2 65.1 83.7 61.4 51.2
Datalab Accurate 70.0 90.3 62.4 83.9 40.8 72.4
Opus 4.8 63.7 89.7 49.8 89.0 71.4 18.7
Azure Document Intelligence 59.6 86.0 1.6 84.9 51.9 73.8
Google Document AI 50.4 55.1 1.4 83.7 50.5 61.3
AWS Textract 47.9 84.6 6.0 74.8 3.7 70.4

Our submitted score is the mean of three full ParseBench runs on a single H100; run-to-run spread was within ±0.2 on every dimension. Every number is reproducible from the submission PR and the settings on the model card.

Throughput

rakedoc-nano is small by design, and small is fast. At 1.2B parameters it saturates a single H100 at about 6 pages/s and fits on any card with 8 GB of memory. It also runs comfortably on an L4, and its small footprint makes it efficient in serverless environments: it scales to zero with minimal cold start.

Every element, located

Parsing for enterprise work is not done when the text is right. It is done when you can point at the page. rakedoc-nano returns a bounding box for every element it emits: paragraphs, headings, list items, table cells, figures, footnotes. That is what lets the Paperwork API redact a clause, sign at an anchor, cite a source region in a retrieval answer, or show an auditor the exact cell a number came from. Most parsing services box images and tables only; we box the text too.

A free tier that does most of the work

Here is the part we are most excited about. Most documents do not need a vision model at all. Native PDFs with a text layer, single-column letters, forms with known templates: our free, rule-based pipeline handles them with no GPU. We already process tens of millions of pages every month. On CloudRaker customer traffic over the last quarter, 75% of documents went through deterministic pipeline. Only the hard 25% (scans, dense tables, multi-column filings) reach OCR and rakedoc-nano.

Most parsing APIs charge for every page. CloudRaker charges for the pages that need a model, at $2 per 1,000 pages rack rate or $0.60 per 1,000 pages for enterprise volume.

Figure 4. Cost per 1,000 pages. Blended rows are an illustration at a 75% free-tier / 25% model mix; your mix depends on your documents. The reference is LlamaParse Cost Effective at $3.75 per 1,000 pages, the cheapest dedicated parsing tier on the ParseBench leaderboard.
Tier Per 1,000 pages Blended at 75 / 25 mix Blended at 50 / 50 mix
Free (rule-based pipeline) $0.00
rakedoc-nano, rack rate $2.00 $0.50 $1.00
rakedoc-nano, enterprise $0.60 $0.15 $0.30
LlamaParse Cost Effective $3.75 $3.75 $3.75

All-in: layout detection, table structure, bounding boxes, free-tier routing, and hosting are included; no per-feature add-ons, no minimums at rack rate. Enterprise pricing applies to committed volume and private deployments. Self-hosted on your own GPU at high utilization, the marginal cost of the model drops below $0.30 per 1,000 pages.

What you can build with it

  • Automated document processing. Extract tables from rate filings, financial statements, and schedules with merged headers intact. This is where the fine-tune moved the needle: perfect table matches rose from 58% to 61% of the benchmark set.
  • Grounded retrieval. Clean Markdown out, tables as HTML, every chunk tied to a box on a page, and nothing leaves your VPC.
  • Agentic paperwork. Parse is the first step of Extract, Compose, Redact, Fill, Redline, and Sign in the Paperwork API and the Agentic Kit for Claude and ChatGPT. Same output contract as the KDL pipeline, so existing integrations work unchanged.

Open by default

rakedoc-nano ships under AGPL-3.0. We publish the weights, the lineage, the exact benchmark commit and settings, and the run-to-run noise we measured (±0.2). We would rather you check our numbers than take our word for them.

Getting started

Parse is one call to the Paperwork API. Send a file, get Markdown, structured JSON, and bounding boxes back. Native PDFs return inline; scans route to rakedoc-nano and take a little longer.

curl -X POST https://api.cloudraker.com/v1/parse \
  -H "Authorization: Bearer $CLOUDRAKER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"file": {"url": "https://example.com/rate-filing.pdf"}}'

Or with the CLI:

paperwork parse --output inline \
  --json '{"file": {"url": "https://example.com/rate-filing.pdf"}}'

Full reference, engines, and polling for long scans: docs.cloudraker.com. Pricing: cloudraker.com/pricing. Grab an API key at signup.cloudraker.com: the forever free tier covers thousands of pages every month.

Send us the documents that break. That is how the next version gets trained.

This is day one of Launch Week. More is coming across the paperwork stack: parsing, extraction, form filling, document generation, and signing.

Weights and model card: huggingface.co/cloudraker/rakedoc-nano.

Model tree. rakedoc-nano is a fine-tune of florin-parser-nano, built on KDL-Frontier-Parser-nano (Qwen2-VL, 1.2B parameters). Both are AGPL-3.0. Serve with vllm serve cloudraker/rakedoc-nano --served-model-name kdl-frontier-parser-nano --trust-remote-code.