Devana

Measured performance

ODIN
in a few numbers.

Devana’s document extraction engine reads a 500-page PDF in 1.5 s, keeps your tables intact and saves the GPU for the pages that need it. Entirely within your infrastructure.

ODIN in four measurements

Measured on November 19, 2026 · ODIN 2.0.32, extraction engine 0.2.0 · one container, 4 CPUs

Speed

500 pages.
1.5 seconds.

ODIN returns the complete content of a 500-page PDF in 1.5 s. Four CPUs are enough.

Up to1,000×faster than the systems compared.

The same 500-page PDF
  • ODIN1.5 s
  • Fastest system compared34 s
  • Slowest system compared24 min 54 s

Page readPage waitingIn real time, at the measured speed: one square per page.

  • PDF · 500 pages

    1.5 s

    341 pages per second

    23× to 1,000× faster

  • Word · 500 pages

    0.04 s

    11,900 pages per second

    3.5× to 160× faster

  • PowerPoint · 100 slides

    0.01 s

    14,300 slides per second

    7.8× to 110× faster

Engine times in fast mode, median of 5 runs. End to end through the service, queue and storage included: 2.3 s for the PDF, 0.86 s for Word, 0.38 s for PowerPoint. 9 systems compared across these three formats, with their default settings.

Tables

Your tables
stay tables.

In a price list, the word “Included” covers three plans. ODIN keeps that span: it renders your tables in HTML, merged cells included.

In Markdown

ServicePlan APlan BPlan C
SupportIncluded
| Support | Included | | |
One cell, one column. The span of the merge is lost.

With ODIN, in HTML

ServicePlan APlan BPlan C
SupportIncluded
<td colspan="3">Included</td>
One cell, three columns. The span is kept.
Merged cells rendered with their exact span
  • 8 of 8

    ODIN, from Word

  • 7 of 8

    ODIN, from PDF

  • 0 of 8

    Systems compared, default output, from PDF

Test document: three tables, 42 cells, 8 of them merged. A merge is counted when the cell appears in the output with its exact span. All 42 cells appear in ODIN’s output.

Fidelity

Your documents.
Word for word.

On a digital document, ODIN reads the text as it is written in the file. On a scanned page, a vision model hosted on your servers takes over.

Digital documentPDF, Word, PowerPoint: the text is read from the file.
  • 100.0%of the words rendered appear in the document
  • 99.9%of the document’s words are rendered
Scanned pageScan or photograph: a vision model transcribes the page.
  • 92.8%fidelity with Mistral Medium 3.5
  • 89.9%fidelity with Mistral Small 3.2

The vision model is a setting: you replace it through configuration.

40 real pages of public documents, compared word for word with their reference text; medians over the 18 pages read by the native parser. Across all 40 pages, with images and tables transcribed by the vision model: 92.0% fidelity. Mistral Small 3.2 and Mistral Medium 3.5: general-purpose vision models, deployable on-premises.

Energy

The GPU,
only when it is needed.

ODIN chooses the path page by page: text is read on CPU, only image pages go to the model on GPU.

2,500×less energy, at minimum, for a page read on CPU.

Page read on CPU
1 drop
55 mJ
Page sent to the model on GPU
2,500 drops
168 J to 5.4 kJ
If a page read on CPU were one drop of water, the same page sent to the model would be at least 2,500.
Extracting a 500-page Word document is like lighting a 10 W LED bulb for:
  • half a secondwith ODIN: 5.2 J

  • 2 h 20 min to 75 hwith an inference on every page: 84 kJ to 2.7 MJ

On a whole documentEnergy saved compared with an inference run on every page with the same model.
  • Word, 500 pages

    99.99%

    of energy saved

    500 pages on CPU · 0 on GPU

    5.2 J with ODIN, versus 84 kJ to 2.7 MJ by inference

  • PowerPoint, 100 slides

    99.99%

    of energy saved

    100 slides on CPU · 0 on GPU

    1.2 J with ODIN, versus 17 kJ to 536 kJ by inference

  • PDF, 500 pages, fast mode

    99.97%

    of energy saved

    500 pages on CPU · 0 on GPU

    27 J with ODIN, versus 84 kJ to 2.7 MJ by inference

  • PDF, 500 pages, default mode

    61%

    of energy saved

    307 pages on CPU · 193 on GPU

    32 kJ to 1 MJ with ODIN, versus 84 kJ to 2.7 MJ by inference

Default mode: 193 of 500 pages contain images or tables sent to the model.

Architecture

Two paths. One infrastructure: yours.

Each page takes the path that suits it. Text is read straight from the file, on CPU; image pages are transcribed by a vision model hosted on your servers. The whole process runs within your infrastructure.

Document

PDF, Word, PowerPoint, image

Page analysis

Page-by-page routing

The page carries its text

Native path

Parser · CPU

The page is an image

Vision path

On-premises model · GPU

Structured content

Page by page, tables in HTML, route of each page

ODIN architecture: each page is routed to the native parser or to the vision model.

Deployment

One container. On your premises.

  • One container, standard CPUs

    The native path runs on standard CPUs. The speed measurements on this page were taken on 4 CPUs.

  • A vision model of your choice

    For image pages: hosted on your servers and replaceable by configuration.

  • Your data stays within your infrastructure

    The native path runs locally; the vision path calls the model hosted on-premises.

  • A single output

    Whatever the document: structured content page by page, tables in HTML, route of each page.

  • A cost proportional to the need

    The GPU is called on for image pages only.

See ODIN at work on your documents.

Our teams present ODIN and its deployment within your infrastructure.

Schedule an ODIN demo