Model reference — Documents · Pictures
Form extraction reads a completed form and returns each answer against the question it answers — name against "name", a ticked box against the option it sits beside, a handwritten date against the field label.
The difficulty is pairing, not reading. A form’s value is in the association between a printed label and the mark someone made next to it, and that association is spatial: labels sit above, beside or inside boxes, and the same form printed by two offices may differ. Models that read words and positions together handle this; text-only extraction returns a jumble. Ticked boxes and handwriting add a second requirement, since neither is text at all until a model interprets it.
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 form at 300 dpi with all fields and ticked boxes extracted.
Vendor: Microsoft
What it does: pairs each printed label with the value written beside it by reading words and positions together. Fitted to a recurring form, it is the most accurate and cheapest option per page.
| 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) | ≈ 1,600 | ≈ 6,000 | ≈ 15,000 |
Vendor: NAVER Clova
What it does: reads the completed form straight from its image, which works where the scan is too poor for reliable character reading. Best when fitted to one form type.
| 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) | ≈ 900 | ≈ 3,200 | ≈ 8,000 |
Vendor: Baidu (PaddlePaddle)
What it does: recovers the form’s ruled structure along with its text, which helps on grid-style forms where answers sit in boxes rather than on lines.
| 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: Microsoft
What it does: reads handwritten entries specifically, which is the part of a form that defeats ordinary character readers. Used as the reading stage for handwritten fields located by another model.
| 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) | ≈ 2,000 | ≈ 8,000 | ≈ 20,000 |
Models in this group take a photograph as input: a completed paper form photographed on a phone or handheld scanner. The three hardware tiers mean the same as above. The sample input here is one photograph of a completed form at about 4 megapixels.
Vendor: Alibaba Cloud
What it does: reads a photographed form and returns each field with its value, including ticked boxes and handwritten notes, following the field list you supply. The most robust option for forms captured in the field.
| 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 (images/hour) | ≈ 400 | ≈ 1,600 | ≈ 900 (72B model, higher accuracy) |
Vendor: OpenGVLab (Shanghai AI Laboratory)
What it does: the same task with strength on dense multi-column forms, such as clinical intake sheets and inspection checklists.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 4090 (2B model) | A100 80 GB (8B model) | 4× H100 80 GB (38B model or larger) |
| VRAM | 12 GB | 40 GB | 320 GB combined |
| vCPUs | 8 | 16 | 48 |
| RAM | 32 GB | 64 GB | 256 GB |
| Server | 1× RTX 4090 24 GB | 1× A100 SXM 80 GB | 4× H100 SXM 80 GB |
| Rate (images/hour) | ≈ 550 | ≈ 1,800 | ≈ 1,200 |
Vendor: Microsoft
What it does: a small vision model that finds the field regions on a photographed form so a reading model only has to look at the parts that matter. Fast enough to run on every capture.
| 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 (images/hour) | ≈ 1,800 | ≈ 6,000 | ≈ 14,000 |
Accuracy depends on whether your forms are a fixed set or an open intake, and how much handwriting they carry. Our consultants review your forms and recommend either a model fitted to each recurring layout or a general vision-language model for varied intake, with the checks that catch a mis-paired field.
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 Form 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.