AI models for Structured data extraction from text
Structured data extraction reads free-written text and fills in the fields you define — order number, delivery date, quantity, site address, fault reported — returning a record that can be written straight into a system without anyone retyping it.
The task is defined by your schema: the list of fields, their types, and which are required. A model is given that schema and made to answer in exactly that shape, which is enforced by the serving software rather than trusted to the model. What separates a working deployment from a demonstration is the handling of fields that are absent or ambiguous — a model must be able to return "not stated" instead of inventing a plausible value.
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 text of about 400 words with a twelve-field record extracted from it.
Llama 3.3 70B
Vendor: Meta
What it does: reads free text and returns a completed record in your schema, handling fields that are implied rather than stated. The choice where a wrong value is costly and volumes are moderate.
Requirement
Minimum
Medium
High
GPU type
2× RTX 4090 (reduced precision)
H100 80 GB
4× H100 80 GB
VRAM
48 GB combined
80 GB
320 GB combined
vCPUs
16
24
64
RAM
64 GB
128 GB
512 GB
Server
2× RTX 4090 24 GB
1× H100 SXM 80 GB
4× H100 SXM 80 GB
Rate (records/hour)
≈ 250
≈ 1,100
≈ 4,500
Qwen2.5 32B
Vendor: Alibaba Cloud
What it does: extracts records from long text and handles many fields at once without degrading, which suits technical reports and long email threads.
Requirement
Minimum
Medium
High
GPU type
RTX 4090 (reduced precision)
L40S 48 GB
2× H100 80 GB
VRAM
22 GB
48 GB
160 GB combined
vCPUs
12
16
48
RAM
48 GB
64 GB
256 GB
Server
1× RTX 4090 24 GB
1× L40S 48 GB
2× H100 SXM 80 GB
Rate (records/hour)
≈ 350
≈ 1,100
≈ 4,200
Mistral Small 3
Vendor: Mistral AI
What it does: a compact model that extracts reliably from short, repetitive text — order emails, service requests — at high volume on one card.
Requirement
Minimum
Medium
High
GPU type
RTX 4090
L40S 48 GB
H100 80 GB
VRAM
24 GB
48 GB
80 GB
vCPUs
12
16
32
RAM
48 GB
64 GB
128 GB
Server
1× RTX 4090 24 GB
1× L40S 48 GB
1× H100 SXM 80 GB
Rate (records/hour)
≈ 800
≈ 2,400
≈ 7,000
Phi-4 14B
Vendor: Microsoft
What it does: a small model that holds up unusually well on numeric fields, so quantities, tolerances and dates survive extraction. Cheap enough to run over an entire archive.
Requirement
Minimum
Medium
High
GPU type
RTX 3090
RTX 4090
H100 80 GB
VRAM
16 GB
24 GB
80 GB
vCPUs
8
12
24
RAM
32 GB
48 GB
128 GB
Server
1× RTX 3090 24 GB
1× RTX 4090 24 GB
1× H100 SXM 80 GB
Rate (records/hour)
≈ 1,000
≈ 3,000
≈ 8,000
GLiNER
Vendor: Urchade Zaratiana and contributors
What it does: pulls out named field values with no training data, from a written list of the fields you want. A quick way to stand up an extraction before deciding whether a larger model is warranted.
Requirement
Minimum
Medium
High
GPU type
RTX 3060
RTX 4090
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 (records/hour)
≈ 3,000
≈ 12,000
≈ 34,000
DeBERTa v3
Vendor: Microsoft
What it does: fitted to your fields from examples, which makes it both the cheapest per record and the most consistent — the right answer once your document types and fields have settled.
Requirement
Minimum
Medium
High
GPU type
RTX 3060
RTX 4090
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 (records/hour)
≈ 8,000
≈ 30,000
≈ 90,000
Choosing between them
The right model depends on how many fields you need, how consistently they appear, and how expensive a wrong value is. Our consultants review your schema and a sample of the text, then recommend a model and the validation rules that catch a bad extraction before it reaches your systems.
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 Structured data extraction from text. 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.