AE3-TL

AE3-TL

Finds every table on a digital PDF page and rebuilds it cell by cell, with merged cells, headers, captions and footers. For teams pulling tables out of reports, filings and papers.

Made by APEXE3HQ 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: AE3-TL is a table extraction model from APEXE3HQ with 32M parameters. It looks at a rendered PDF page and the page's own words, then returns each table as a grid of cells. Headers, merged cells, cell text, captions and footers all come back linked to the right table.

Who it suits: Teams that pull tables out of digital PDFs, such as reports, filings, papers and statements, and need the cell text to match the document exactly. The card says it runs on a single GPU or on a CPU through ONNX.

Title: Who uses it, and how
Industry | Who uses it | How they use it | Example
Finance | Analysts | Pull tables out of filings and statements | Turn a balance sheet page into rows and columns
Research | Researchers | Extract result tables from papers | Collect the tables from a batch of PDF papers
Data engineering | Pipeline builders | Run extraction over large document sets | Process a backlog of reports page by page
Document software | Developers | Add table recognition to a PDF tool | Show each detected table with its caption attached

Skip it if your PDFs are scans without a text layer, because the cell text comes from the PDF's own words rather than from reading the image.

Where to use it:

• Python or any program that runs ONNX models
• A single GPU, or a CPU for smaller jobs
• Windows, macOS or Linux for downloading the files
• Document pipelines that already render PDF pages and read their words

Use case | Who | What they get back
Table extraction | Analysts | Each table rebuilt cell for cell, with headers and merged cells
Caption linking | Document teams | Captions and footers tied to the right table on a busy page
Bulk processing | Engineers | Tables from many pages at the card's reported speed

How to set it up: You need Python and pip to download, plus a way to run ONNX models. You also need to turn each PDF page into an image and read its words with positions.

• Python and pip
• An ONNX runtime
• A step that renders PDF pages and reads their text layer

1. Install the Hugging Face hub library.
2. Download the repository.
3. Feed it a rendered page and the page's words through your ONNX runtime.

pip install -U huggingface_hub
hf download APEXE3HQ/ae3-tl

How to use it: Give the model two inputs per page, the rendered image and the words with their positions. It returns every table on the page. The card points to its code repository for the full run steps.

Tool | What it does
Rendered page image | The first input, where the model reads the table structure
Words with positions | The second input, taken from the PDF's text layer
Grid output | Each table's cells, headers, text, caption and footer

Example: You give it a page from an annual report with two tables and their captions. It returns each table as a grid of cells with the merged header cells kept, and links each caption to the table below it.

For newcomers and experts: If you are new, think of it as a tool that sees where a table's lines and cells are and copies the words from the PDF into them. Start with one page that has a single table.

For people in the field, the structure comes from the page image while the text comes from the words the PDF already holds, which avoids reading mistakes. On the PubTables-v2 page-level test, 73.7 percent of tables came out as an exact match, against POTATR at 64.5 and Gemini 3.1 Pro at 35.2.

Render the PDF page as an image
Read the page's words and positions
Model finds each table's structure
Cells, captions and footers come back

Good to know:

Title: Tables reproduced exactly, percent of 9,047 test tables
AE3-TL: 73.7%
POTATR: 64.5%
Gemini 3.1 Pro: 35.2%

• License: Apache 2.0, including commercial use as the card states
• Cost to run: none named, though the card's own figure is about $30 per million pages on a rented A10 GPU
• Data it reaches: it downloads files from Hugging Face, then reads the page images and words you give it
• Not for: scanned PDFs with no text layer
• Speed reached about 80,808 pages per hour when measured on a single A10 GPU

On Hugging Face, 13 November 2026

• Task: object detection
• Library: onnx
• Downloads in the last 30 days: 16
• Likes: 6
• License: apache-2.0
• Languages: en
• Tags: onnx, coreml, table-extraction, table-structure-recognition, document-ai, d-fine, object-detection

Links

• Model: https://huggingface.co/APEXE3HQ/ae3-tl
• Files: https://huggingface.co/APEXE3HQ/ae3-tl/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download APEXE3HQ/ae3-tl

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