APEXE3

Small, specialised models that beat much larger ones, and that organisations can run privately at scale.

For seven years APEXE3 has worked with large asset managers. Across all of them, one need keeps coming up: getting accurate, structured data out of millions of documents without sending them anywhere. We build compact models for exactly that: accurate enough to trust, cheap enough to run on every page, and small enough to deploy inside a client's own infrastructure.

Our first model: AE3-TL, page-level table extraction

AE3-TL finds every table on a PDF page and rebuilds it cell for cell, with captions and footers attached to the right table.

• State of the art on PubTables-v2: first on all six metrics of the page-level benchmark.
• 73.7% of tables reproduced exactly, against 64.5% for the previous best (POTATR) and 35.2% for Gemini 3.1 Pro.
• The harder the page, the bigger the lead: 59% exact on pages with more than 250 cells, against 40%.
• 32M parameters: about 80,000 pages an hour on a single GPU, roughly $30 per million pages.
• Free and open under Apache 2.0, including for commercial use.

Model · Paper · Code

What's next

We are extending the same approach to the rest of the document:

• Charts: recovering the underlying data from plotted figures
• Infographics: reading layouts that mix text, numbers and graphics

Work with us

We help teams run our models privately at scale, tune them on their own document types, and integrate them into research and data pipelines.

labs@apexe3.com · apexe3.com