Relation extraction records how the things in a document are connected: this company acquired that one, in this year, for this amount; this person signed on behalf of that entity; this component supersedes that part number.
Where entity recognition gives you a list of names, relation extraction gives you statements — subject, relationship, object — that can be loaded into a database or a knowledge graph and queried. That makes it the foundation of due-diligence research, supply-chain mapping and regulatory monitoring. The hard part is not finding the facts but keeping them consistent: the same company written three ways must resolve to one record, and a fact must carry the sentence it came from so it can be checked.
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 of about 500 words with all relations extracted.
Llama 3.3 70B
Vendor: Meta
What it does: reads a page and writes out the facts it states as structured statements, following your own schema of relationship types. The most accurate option where facts span several sentences and require inference.
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 (pages/hour)
≈ 200
≈ 900
≈ 4,000
Qwen2.5 32B
Vendor: Alibaba Cloud
What it does: extracts relations from very long documents in one pass, so a whole annual report can be processed without splitting and losing cross-references.
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 (pages/hour)
≈ 300
≈ 900
≈ 3,600
Mistral Small 3
Vendor: Mistral AI
What it does: a compact model for high-volume relation extraction on a single card, suited to monitoring a daily feed of filings or news rather than deep analysis of a few documents.
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 (pages/hour)
≈ 500
≈ 1,500
≈ 4,500
GLiNER (relation mode)
Vendor: Urchade Zaratiana and contributors
What it does: finds pairs of entities in a stated relationship — "supplier of", "subsidiary of" — from a written description, with no training data. Very fast and a sound first pass before a larger model looks at the difficult pages.
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 (pages/hour)
≈ 2,400
≈ 9,000
≈ 26,000
REBEL / DeBERTa relation classifier
Vendor: Microsoft
What it does: a small purpose-built model that classifies the relationship between two already-identified entities. Cheap, quick and, once fitted to your relationship types, accurate on well-formed sentences.
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 (pages/hour)
≈ 10,000
≈ 40,000
≈ 110,000
BGE-M3
Vendor: Beijing Academy of Artificial Intelligence
What it does: fingerprints extracted facts so that the same statement phrased differently in two documents is recognised as one fact rather than two, which is what keeps the resulting database usable.
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)
≈ 6,000
≈ 30,000
≈ 90,000
Choosing between them
This is more pipeline than model: extraction, then resolving duplicate names, then storing the facts with their sources. Our consultants review your documents and the questions you want to ask of the result, then recommend the models and the storage that fit.
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 Fact / relation 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.