Model reference — Documents
This service converts a document into a machine-readable record of what is on each page: headings, paragraphs, lists, tables, figures, captions and footnotes, each with its position and its place in the reading order.
It is the groundwork for everything else. Search, question answering, extraction and archiving all work better on a structured document than on a wall of extracted text, because structure carries meaning — a number under a heading belongs to that heading. Models divide into pipelines, which detect layout and then read each region, and single vision-language models, which read the page and emit the structure directly. Pipelines are faster and more predictable; vision-language models cope better with unusual layouts.
Models in this group take text or whole documents as input: plain text, 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. Use these rates for initial sizing. Before production, benchmark your own data to validate accuracy, latency, throughput and cost. The sample input here is one A4 page at 300 dpi converted to a structured record.
Vendor: IBM
What it does: converts a whole document into a structured record with headings, tables, lists and reading order preserved, and exports it as JSON or Markdown. A dependable general-purpose converter for business documents.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | No GPU — CPU only | RTX 4090 | A100 80 GB |
| VRAM | — | 24 GB | 80 GB |
| vCPUs | 8 | 12 | 24 |
| RAM | 16 GB | 32 GB | 64 GB |
| Server | CPU instance, 8 vCPU | 1× RTX 4090 24 GB | 1× A100 SXM 80 GB |
| Rate (pages/hour) | ≈ 250 | ≈ 2,400 | ≈ 5,000 |
Vendor: OpenDataLab (Shanghai AI Laboratory)
What it does: converts PDFs in bulk while keeping tables, mathematical formulas and figures intact. Built for turning a large archive into machine-readable data in one job.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3090 | RTX 4090 | 2× A100 80 GB |
| VRAM | 12 GB | 24 GB | 160 GB combined |
| vCPUs | 8 | 16 | 32 |
| RAM | 32 GB | 48 GB | 128 GB |
| Server | 1× RTX 3090 24 GB | 1× RTX 4090 24 GB | 2× A100 SXM 80 GB |
| Rate (pages/hour) | ≈ 500 | ≈ 1,800 | ≈ 5,500 |
Vendor: Baidu (PaddlePaddle)
What it does: runs layout analysis, table structure and character reading in a single pass, so one run returns the full page structure. Strongest on English and Chinese text.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3060 | RTX 4090 | L40S 48 GB |
| VRAM | 8 GB | 24 GB | 48 GB |
| vCPUs | 6 | 12 | 24 |
| RAM | 16 GB | 32 GB | 64 GB |
| Server | 1× RTX 3060 12 GB | 1× RTX 4090 24 GB | 1× L40S 48 GB |
| Rate (pages/hour) | ≈ 900 | ≈ 3,600 | ≈ 6,500 |
Vendor: Datalab
What it does: detects layout, works out reading order and reads the characters in roughly ninety languages. A practical choice for multilingual archives and uneven scans.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3060 | RTX 4090 | A100 80 GB |
| VRAM | 8 GB | 24 GB | 80 GB |
| vCPUs | 6 | 12 | 24 |
| RAM | 16 GB | 32 GB | 64 GB |
| Server | 1× RTX 3060 12 GB | 1× RTX 4090 24 GB | 1× A100 SXM 80 GB |
| Rate (pages/hour) | ≈ 700 | ≈ 3,000 | ≈ 6,000 |
Vendor: Alibaba Cloud
What it does: reads the page as an image and writes out its structure directly, coping with layouts a pipeline mis-parses — magazine pages, engineering drawings, annotated forms. Slower and dearer per page.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 4090 (7B, reduced precision) | A100 80 GB (7B, full precision) | 2× H100 80 GB (72B model) |
| VRAM | 16 GB | 80 GB | 160 GB combined |
| vCPUs | 8 | 16 | 32 |
| RAM | 32 GB | 64 GB | 200 GB |
| Server | 1× RTX 4090 24 GB | 1× A100 SXM 80 GB | 2× H100 SXM 80 GB |
| Rate (pages/hour) | ≈ 400 | ≈ 1,600 | ≈ 900 (72B model, higher accuracy) |
Vendor: StepFun
What it does: reads a page in one step and outputs tables, formulas and text as structured markup. Small enough to run on a modest card, which suits deployment at a branch site.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3060 | RTX 4090 | A100 80 GB |
| VRAM | 8 GB | 24 GB | 80 GB |
| vCPUs | 6 | 12 | 24 |
| RAM | 16 GB | 32 GB | 64 GB |
| Server | 1× RTX 3060 12 GB | 1× RTX 4090 24 GB | 1× A100 SXM 80 GB |
| Rate (pages/hour) | ≈ 600 | ≈ 2,200 | ≈ 4,500 |
Vendor: community-maintained open-source project
What it does: the long-established character reader. It returns words and their positions but no document structure, so it is used as the reading stage under a layout model rather than on its own.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | No GPU required | No GPU required | No GPU required |
| VRAM | — | — | — |
| vCPUs | 2 | 8 | 32 |
| RAM | 4 GB | 16 GB | 64 GB |
| Server | CPU instance, 2 vCPU | CPU instance, 8 vCPU | CPU instance, 32 vCPU |
| Rate (pages/hour) | ≈ 2,000 | ≈ 8,000 | ≈ 30,000 |
Which route fits depends on how uniform your documents are and how much structure you need preserved. Our consultants review a sample of your pages, compare a pipeline against a vision-language model on them, and recommend the one that gives the structure your downstream systems need at the throughput you require.
At the start of a project we may run a short proof of concept on a sample of your own data, 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.
Share a representative sample, expected volume, latency target and deployment location for PDF/document to structured JSON. 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.