Introducing decision-machine-1: a language model for decisions, not prose

rakedoc-nano Available now

Every element, located.

A 1.2-billion-parameter vision model that turns PDFs, scans and filings into clean markdown and structured tables, with a bounding box on every block it emits. Not only the pictures. The text too.

Weights on Hugging FaceRead the Parse docs
ParseBench overall
77.2
First among all vision-language models, fifth overall.
Per 1,000 pages
$0.60
At enterprise volume, and only on the pages that need a model.
Parameters
1.2B
Runs on any GPU with 8 GB. About 6 pages a second on one H100.
Weights
AGPL-3.0
Published on Hugging Face. Run it inside your own perimeter.

01The business case

rakedoc-nano

Reading the document stops being the expensive part.

Every workflow that touches paperwork starts by reading it. Most parsing bills charge for that on every page, whether the page needed a model or not.

Three pages in four are free
Native PDFs, letters and known forms go through a rule-based pipeline with no GPU. You pay for the hard quarter, not the whole stack.
Checkable, not claimed
77.2 on the public ParseBench leaderboard, first among all vision-language models. The run, the settings and the weights are all published.
Point at the page
A box on every element, table cells included. That is what lets you redact a clause, cite a source, or show an auditor the exact cell.
Yours to run
Open weights under AGPL-3.0 on a card with 8 GB. Documents never have to leave your perimeter to be read.

02The model

rakedoc-nano

Built for the documents that actually turn up.

The pages that matter in insurance, healthcare, legal and financial services are the ones with a thirty-column rate table and a three-level header. That is where we aimed it.

  • A dense table with merged headers
    Generic OCR Text in reading order, the grid flattened, the header hierarchy gone.
    rakedoc-nano 86.4 on the Tables dimension, ahead of every dedicated parsing API but one, with the structure intact.
  • Knowing where a number came from
    Generic OCR Boxes on the images and the tables, if you are lucky. Nothing on the text.
    rakedoc-nano A box on every element it emits: paragraphs, headings, list items, table cells, figures, footnotes.
  • What a page costs
    Generic OCR Every page billed at the model rate, including the ones with a perfectly good text layer.
    rakedoc-nano Three in four pages never reach a model. The rest are $0.60 per thousand at enterprise volume.
  • Where the documents go
    Generic OCR A closed endpoint, a black-box score, and your filings on someone else's hardware.
    rakedoc-nano Open weights, a published benchmark run, and vLLM on your own GPU if that is what compliance needs.

03Specification

rakedoc-nano

Everything it is, on one page.

One model, one licence, one price on the pages that need it. The lineage is published because a parser you cannot inspect is a parser you have to trust.

Model

Name
rakedoc-nano
Category
Vision-language parser Built by CloudRaker, and the engine behind the Parse capability.
Task
Document parsing Markdown, tables as structure, and a bounding box on every element.
Parameters
1.2 billion
Lineage
florin-parser-nano, on KDL-Frontier-Parser-nano (Qwen2-VL) A fine-tune. Both parents are AGPL-3.0.

Running it

Weights
huggingface.co/cloudraker/rakedoc-nano
Licence
AGPL-3.0 Run it, inspect it, fine-tune it.
Self-hosted
vllm serve cloudraker/rakedoc-nano Any GPU with 8 GB. It runs comfortably on an L4 and scales to zero.
Throughput
About 6 pages a second Saturating a single H100.
Hosted
Paperwork API and your CloudRaker Workspace

Price

Rule-based tier
$0 Where three pages in four are handled, with no GPU involved.
Rack rate
$2.00 per 1,000 pages Charged only on the pages that reach a model. No minimums.
Enterprise
$0.60 per 1,000 pages Committed volume and private deployments.

All in. Layout detection, table structure, bounding boxes, free-tier routing and hosting are included, with no per-feature add-ons.

See the model card

04Benchmarks

rakedoc-nano

The parser is ours. So are the weights.

rakedoc-nano is the model that reads the hard quarter. It is first among all vision-language models on ParseBench (proprietary frontier models included), and it is published on Hugging Face under AGPL-3.0, so you can check that yourself.

  • 77.2 ParseBench overall First among all VLMs, fifth overall
  • 1.2 B Parameters Runs on any GPU with 8 GB, ~6 pages/s on one H100
  • $0.60 Per 1,000 pages At enterprise volume; $2 rack rate, and only on pages that need a model
  • AGPL-3.0 Open weights Run it, inspect it, fine-tune it, keep documents inside your perimeter

ParseBench overall score, leaderboard, including our submitted run

  1. LlamaParse Cost Effective 80.6
  2. rakedoc-nano ours, open weights 77.2
  3. florin-parser-nano open 76.7
  4. Gemini 3 Flash 75.1
  5. Reducto (Agentic) 73.0
  6. Fable 5 70.8
  7. Datalab Accurate 70.0
  8. Opus 4.8 63.7
  9. Azure Document Intelligence 59.6
  10. Google Document AI 50.4
  11. AWS Textract 47.9

rakedoc-nano, per dimension

  • 86.4 Tables Ahead of every dedicated parsing API but one
  • 88.8 Content faithfulness Within 3.5 points of the best on the bench
  • 71.7 Semantic formatting Sixth overall, third among VLMs
  • 74.3 Grounding The boxes, scored
  • 64.9 Charts The dimension we are least proud of

Every number here comes from the public ParseBench leaderboard, including our own submitted run. Our score is the mean of three full runs on a single H100; run-to-run spread stayed within ±0.2 on every dimension. Models with no published price are on the chart anyway. We would rather you check the numbers than take our word for them.

Weights and model cardSee the benchmark

05Grounding

rakedoc-nano

Parsing 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, and most parsing services box the images and the tables only. We box the text too, down to the individual table cell.

That is not a detail. It is the difference between a wall of markdown and a document your software can act on, and it is what the rest of the Paperwork API is built on top of.

  • Redact a clause at its coordinates, not by a string match that may miss.
  • Sign at an anchor that is on the page, wherever the layout put it.
  • Cite the source region in a retrieval answer, so the reader can check it.
  • Show an auditor the exact cell a number came from.

06Pricing

rakedoc-nano

Most pages do not need a model at all.

Native PDFs with a text layer, single-column letters, forms with a known template. Our rule-based pipeline handles them with no GPU and no charge. Only the hard remainder reaches rakedoc-nano.

75 %

of CloudRaker customer documents went through the free, deterministic pipeline last quarter. Most parsing APIs would have billed for every one of them.

Rule-based tier, per 1,000 pages
$0.00
rakedoc-nano, rack rate
$2.00
rakedoc-nano, enterprise volume
$0.60
Blended at a 75 / 25 mix, enterprise
$0.15
See CloudRaker pricing

07Integration

rakedoc-nano

One call, and the page comes back readable.

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"}}'
import { CloudRakerClient } from "@cloudraker/api";

const client = new CloudRakerClient({ token: "YOUR_API_KEY" });

const run = await client.parse({
  body: {
    file: "{ url: \"https://www.irs.gov/pub/irs-pdf/fw9.pdf\", name: \"w9.pdf\" },"
  },
  output: "inline",
});

console.log(run.output?.markdown); // clean markdown, straight from the PDF
from cloudraker import V1ParseBodyFileName
from cloudraker.client import CloudRaker

client = CloudRaker(token="YOUR_API_KEY")

run = client.parse(
    file=V1ParseBodyFileName(url="https://www.irs.gov/pub/irs-pdf/fw9.pdf", name="w9.pdf"),
    output="inline",
)

print(run.output.markdown)  # clean markdown, straight from the PDF

08Questions

rakedoc-nano

Important questions, fast answers.

Is 77.2 actually good?

It is first among every vision-language model on ParseBench, proprietary frontier models included, and fifth overall behind three managed LlamaParse tiers and one model with no published price. At a fraction of the cost of everything above it.

Where is it weakest?

Charts, at 64.9. If your documents are mostly chart-heavy decks rather than filings and forms, benchmark it against your own pages before committing.

Can I run it myself?

Yes. The weights are on Hugging Face under AGPL-3.0 and it serves with vLLM on any GPU with 8 GB of memory. At high utilisation on your own hardware the marginal cost drops below $0.30 per 1,000 pages.

How do I know which pages were billed?

Only pages that reach a model are charged. Every run reports the route it took, so the split between the free tier and rakedoc-nano is visible per document rather than estimated.

Does it handle audio?

Not itself. Recordings route to rakeaudio-asr, which transcribes with speaker diarization and returns the same shape, because a conversation is a document too.

Where to go next

rakedoc-nano is one of three.

Same infrastructure under each one. Pick the model that fits the input you actually have.

decision-machine-1
Sub-second typed decisions over text. Yes/no, classify, rate, answer, extract, entities and verify, each with a probability attached and nothing generated.
See decision-machine-1
rakeaudio-asr
Speech to text with speaker diarization, because a conversation is a document too. Recordings come back in the same shape a page does.
See rakeaudio-asr
Get started

Send us the documents that break.

The free tier covers thousands of pages a month. Run it on the pages your current parser gets wrong, and tell us where it fails. That is how the next version gets trained.

Start for freeTalk to us