Model reference — Documents · Pictures
Document classification decides what a document is — invoice, contract, passport, delivery note, medical referral — so that it can be routed to the right team, filed in the right place and processed by the right extraction rules without anyone opening it.
It is usually the first step in a document pipeline, and it can be done in two ways. A text classifier reads the words, which works when the wording is distinctive. An image classifier reads the page as a picture, which works when the shape of the page gives it away — a form’s ruling lines, a letterhead, a stamp — and keeps working on scans too poor for reliable text extraction. Models that read words and layout together are the most accurate on mixed intake.
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 assigned to one document type.
Vendor: Microsoft
What it does: reads the words on a page together with their positions, so it distinguishes document types by both wording and layout. The strongest general choice for a varied document intake.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3060 | RTX 4090 | 1× 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) | ≈ 2,000 | ≈ 8,000 | ≈ 20,000 |
Vendor: Microsoft
What it does: classifies on the extracted text alone. The cheapest and fastest route when your documents are born-digital and their wording clearly identifies them.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3060 | RTX 4090 | 1× A100 80 GB |
| VRAM | 6 GB | 24 GB | 80 GB |
| vCPUs | 4 | 8 | 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) | ≈ 20,000 | ≈ 90,000 | ≈ 260,000 |
Vendor: NAVER Clova
What it does: classifies a page straight from its image with no OCR step, which removes a source of error on faxes and poor photocopies.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3060 | RTX 4090 | 1× 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) | ≈ 1,800 | ≈ 6,500 | ≈ 16,000 |
Vendor: Alibaba Cloud
What it does: classifies from a written description of each document type rather than from labelled examples, which lets a new type be added the same day it appears. Slower and dearer per page than a fitted classifier.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 4090 (reduced precision) | 1× A100 80 GB | 2× A100 80 GB |
| VRAM | 16 GB | 80 GB | 160 GB combined |
| vCPUs | 8 | 16 | 32 |
| RAM | 32 GB | 64 GB | 128 GB |
| Server | 1× RTX 4090 24 GB | 1× A100 SXM 80 GB | 2× A100 SXM 80 GB |
| Rate (pages/hour) | ≈ 400 | ≈ 1,600 | ≈ 3,000 |
Models in this group take a single image as input: a photograph of a page, a screenshot or a scanned form. The three hardware tiers mean the same as above. The sample input here is one page image at about 2 megapixels assigned to one document type.
Vendor: Microsoft
What it does: classifies a page purely by its appearance, trained on millions of scanned business documents. Very fast, and effective where each document type has a recognisable shape.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3060 | RTX 4090 | 1× A100 80 GB |
| VRAM | 6 GB | 24 GB | 80 GB |
| vCPUs | 4 | 8 | 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 (images/hour) | ≈ 12,000 | ≈ 50,000 | ≈ 140,000 |
Vendor: Google
What it does: a general image classifier fitted to your document types. Small, quick and easy to retrain, which suits a fixed set of forms photographed in a controlled way.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3060 | RTX 4090 | 1× A100 80 GB |
| VRAM | 6 GB | 24 GB | 80 GB |
| vCPUs | 4 | 8 | 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 (images/hour) | ≈ 20,000 | ≈ 80,000 | ≈ 220,000 |
Vendor: Microsoft
What it does: the long-established image classifier. Less accurate than newer models but exceptionally well understood and supported on almost any hardware, which matters for a deployment that must run at a remote site for years.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3060 | RTX 4090 | 1× A100 80 GB |
| VRAM | 4 GB | 24 GB | 80 GB |
| vCPUs | 4 | 8 | 24 |
| RAM | 8 GB | 32 GB | 64 GB |
| Server | 1× RTX 3060 12 GB | 1× RTX 4090 24 GB | 1× A100 SXM 80 GB |
| Rate (images/hour) | ≈ 30,000 | ≈ 120,000 | ≈ 320,000 |
Accuracy depends on how similar your document types look and how clean the scans are. Our consultants review a sample of your intake, measure where the confusions fall, and recommend a text, image or combined classifier along with the confidence threshold at which a document should go to a person instead.
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 Document classification. 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.