Skip to main content

Model reference — Documents

AI models for Table extraction

Table extraction takes a page — a photocopy, a scan, or a PDF produced directly by software — and returns the tables printed on it as rows and columns. Cell text, merged headers and the order in which cells should be read are all preserved, so the result opens in a spreadsheet or loads straight into a database instead of being typed out by hand.

Two families of model do this work. Layout-detection models locate the table on the page and rebuild its grid of cells, then hand the cells to an OCR (optical character recognition) step that reads the characters. Vision-language models skip that split: they look at the whole page image and write the table out directly as structured text. Most working pipelines use one of each. The models below cover both families, at every scale we deploy.

Table extraction service All document models

Input type — Documents

Models in this group take whole documents as input: PDFs, scanned pages and office files. Each specification table gives three hardware tiers — Minimum, the smallest setup on which the model runs correctly; Medium, the usual production configuration; and High, a configuration sized for peak volume. The rate is the number of sample inputs processed per hour on that hardware. The sample input here is one A4 page scanned at 300 dpi carrying a single table. Use these rates for initial sizing. Before production, benchmark your own data to validate accuracy, latency, throughput and cost.

Table Transformer (DETR)

Vendor: Microsoft

What it does: locates each table on a page and reconstructs its grid of rows, columns and merged cells. Use it for tables with printed ruling lines — financial statements, regulatory returns, printed reports — at high volume. It returns the grid only, so it is paired with an OCR step that reads the text inside each cell.

RequirementMinimumMediumHigh
GPU typeRTX 3060RTX 4090A100 80 GB
VRAM6 GB24 GB80 GB
vCPUs4816
RAM16 GB32 GB64 GB
Server1× RTX 3060 12 GB1× RTX 4090 24 GB1× A100 SXM 80 GB
Rate (pages/hour)≈ 1,800≈ 7,000≈ 14,000

PP-StructureV3

Vendor: Baidu (PaddlePaddle)

What it does: page layout, table structure and character reading in a single pass, so one run returns both the tables and the paragraphs around them. Accuracy is strongest on English and Chinese text. A sound default when the whole page needs converting, not just the tables.

RequirementMinimumMediumHigh
GPU typeRTX 3060RTX 4090L40S
VRAM8 GB24 GB48 GB
vCPUs61224
RAM16 GB32 GB64 GB
Server1× RTX 3060 12 GB1× RTX 4090 24 GB1× L40S 48 GB
Rate (pages/hour)≈ 900≈ 3,600≈ 6,500

Docling / TableFormer

Vendor: IBM

What it does: converts an entire document into structured output and recovers tables whose headers are merged or stacked several rows deep. Suited to financial and scientific reports, where a figure means nothing unless the header above it is attached correctly.

RequirementMinimumMediumHigh
GPU typeNo GPU — CPU onlyRTX 4090A100 80 GB
VRAM24 GB80 GB
vCPUs81224
RAM16 GB32 GB64 GB
ServerCPU instance, 8 vCPU1× RTX 4090 24 GB1× A100 SXM 80 GB
Rate (pages/hour)≈ 250≈ 2,400≈ 5,000

Surya

Vendor: Datalab

What it does: detects layout, works out the order in which blocks should be read, reads the characters and recovers table structure, in roughly ninety languages. A practical choice for multilingual archives and for scans of uneven quality.

RequirementMinimumMediumHigh
GPU typeRTX 3060RTX 4090A100 80 GB
VRAM8 GB24 GB80 GB
vCPUs61224
RAM16 GB32 GB64 GB
Server1× RTX 3060 12 GB1× RTX 4090 24 GB1× A100 SXM 80 GB
Rate (pages/hour)≈ 700≈ 3,000≈ 6,000

MinerU

Vendor: OpenDataLab (Shanghai AI Laboratory)

What it does: converts PDFs in bulk while keeping tables, mathematical formulas and figures intact. Built for one-off jobs that turn a large document archive into machine-readable data.

RequirementMinimumMediumHigh
GPU typeRTX 3090RTX 40902× A100 80 GB
VRAM12 GB24 GB160 GB combined
vCPUs81632
RAM32 GB48 GB128 GB
Server1× RTX 3090 24 GB1× RTX 4090 24 GB2× A100 SXM 80 GB
Rate (pages/hour)≈ 500≈ 1,800≈ 5,500

Camelot / pdfplumber

Vendor: community-maintained open-source projects

What it does: reads the ruling lines and text positions already stored inside a PDF that was produced by software rather than scanned, and rebuilds the table from them. No neural model and no GPU are involved, so results are exact and repeatable — but it cannot read scans or photographs, which carry no text layer.

RequirementMinimumMediumHigh
GPU typeNo GPU requiredNo GPU requiredNo GPU required
VRAM
vCPUs2832
RAM4 GB16 GB64 GB
ServerCPU instance, 2 vCPUCPU instance, 8 vCPUCPU instance, 32 vCPU
Rate (pages/hour)≈ 3,000≈ 12,000≈ 45,000

Azure Document Intelligence

Vendor: Microsoft

What it does: returns tables, labelled field values (such as an invoice number next to its label) and ticked checkboxes from a page. It ships as a container that runs inside your own cluster, so documents never leave your network.

RequirementMinimumMediumHigh
GPU typeNoneNoneRTX 4090
VRAM24 GB
vCPUs2816
RAM4 GB16 GB32 GB
ServerNot applicableNot applicable1× RTX 4090 24 GB (container)
Rate (pages/hour)≈ 900≈ 9,000≈ 4,000

Amazon Textract

Vendor: Amazon Web Services

What it does: analyses tables and forms, and accepts long documents as background jobs that report back when finished rather than making the caller wait. Usually chosen when the rest of the pipeline already runs on Amazon Web Services.

RequirementMinimumMediumHigh
GPU typeNo GPU requiredNo GPU requiredNo GPU required
VRAM
vCPUs2816
RAM4 GB16 GB32 GB
ServerNot applicableNot applicableNot applicable
Rate (pages/hour)≈ 600 (one page per request)≈ 10,000 (background batch jobs)≈ 36,000 (with raised account limits)

Google Document AI

Vendor: Google Cloud

What it does: reads page layout and forms and returns the tables found. It can also be trained on a document type you receive repeatedly — one supplier’s invoice, one agency’s form — to pull named fields out of it reliably.

RequirementMinimumMediumHigh
GPU typeNo GPU requiredNo GPU requiredNo GPU required
VRAM
vCPUs2816
RAM4 GB16 GB32 GB
ServerNot applicableNot applicableNot applicable
Rate (pages/hour)≈ 720 (one page per request)≈ 12,000 (background batch jobs)≈ 40,000 (with raised account limits)

Mistral OCR

Vendor: Mistral AI

What it does: reads a document and returns it as Markdown with the tables intact, coping with handwriting and dense multi-column layouts that defeat simpler extractors.

RequirementMinimumMediumHigh
GPU typeNo GPU requiredNo GPU requiredNo GPU required
VRAM
vCPUs2816
RAM4 GB16 GB32 GB
ServerNot applicableNot applicableNot applicable
Rate (pages/hour)≈ 1,200≈ 20,000 (background batch jobs)≈ 60,000 (background jobs, raised limits)

Input type — Pictures

Models in this group take a single image as input: a phone photograph of a paper form, a screenshot of a slide, a picture of a grid drawn on a whiteboard. They read the image directly and return the table as structured text. The three hardware tiers mean the same as above, and the sample input is one photograph of a page carrying a single table.

Qwen2.5-VL 7B / 72B

Vendor: Alibaba Cloud

What it does: looks at a photograph or scan of a table and writes it out as structured text. It tolerates tilted pages, poor lighting and handwriting better than layout-detection models, and it follows a written instruction about the output you want — for example which columns to keep, or how to format dates. The 7B and 72B labels are model sizes in billions of parameters: the larger one is more accurate and slower.

RequirementMinimumMediumHigh
GPU typeRTX 4090 (7B, reduced precision)A100 80 GB (7B, full precision)2× H100 80 GB (72B model)
VRAM16 GB80 GB160 GB combined
vCPUs81632
RAM32 GB64 GB200 GB
Server1× RTX 4090 24 GB1× A100 SXM 80 GB2× H100 SXM 80 GB
Rate (images/hour)≈ 400≈ 1,600≈ 900 (72B model, higher accuracy)

InternVL 3

Vendor: OpenGVLab (Shanghai AI Laboratory)

What it does: reads charts and tables from page images and returns them as structured text. It comes in several sizes; the 8-billion-parameter version is the usual balance of accuracy and speed for on-premise deployment, while larger versions read dense grids more accurately at lower throughput.

RequirementMinimumMediumHigh
GPU typeRTX 4090 (2B model)A100 80 GB (8B model)4× H100 80 GB (38B model or larger)
VRAM12 GB40 GB320 GB combined
vCPUs81648
RAM32 GB64 GB256 GB
Server1× RTX 4090 24 GB1× A100 SXM 80 GB4× H100 SXM 80 GB
Rate (images/hour)≈ 550≈ 1,800≈ 1,200

GOT-OCR 2.0

Vendor: StepFun

What it does: reads a page image in one step and outputs tables, mathematical formulas and even sheet music as structured markup. It is small enough to run on a modest graphics card, which makes it suitable for deployment at a branch office or on equipment at the edge of the network.

RequirementMinimumMediumHigh
GPU typeRTX 3060RTX 4090A100 80 GB
VRAM8 GB24 GB80 GB
vCPUs61224
RAM16 GB32 GB64 GB
Server1× RTX 3060 12 GB1× RTX 4090 24 GB1× A100 SXM 80 GB
Rate (images/hour)≈ 600≈ 2,200≈ 4,500

Choosing between them

No single model wins on every document set. Our consultants review your material, the accuracy you need and the volume you expect, then recommend the model and pipeline that fit: a rule-based route for PDFs produced by software, a layout-detection model for ruled scans at volume, a vision-language model for photographs, handwriting and irregular grids.

At the start of a project we may run a short proof of concept on a sample of your own documents, measuring the accuracy and the throughput the model actually achieves on your material. That replaces the estimates on this page with real figures, so the cost and the schedule for the full engagement are known before it is committed.

Table extraction service AI models for documents Pricing

From benchmark to production

Share a representative sample, expected volume, latency target and deployment location for Table extraction. We benchmark the shortlisted models, recommend the lowest-cost GPU configuration that meets the target, and scale it from pilot capacity to a dedicated production cluster.