Text summarization reduces a long document to its substance at a length you choose — a paragraph, a page, a set of bullet points — while keeping the facts, figures and names that matter and dropping the rest.
Two approaches exist. Extractive models select the most important sentences from the original and return them unchanged, which guarantees nothing is invented but reads awkwardly. Abstractive models — today’s large language models — rewrite the substance in new sentences, which reads naturally but requires checks that no figure has drifted. The length a model can read at once matters as much as its quality: a model with a short reading window must have long documents split and summarised in stages.
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 summarised to roughly 100 words.
Llama 3.3 70B
Vendor: Meta
What it does: rewrites long documents into fluent summaries and follows instructions about length, tone and what to keep. The most reliable choice when the summary will be read by a customer or an executive.
Requirement
Minimum
Medium
High
GPU type
2× RTX 4090 (reduced precision)
1× 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)
≈ 300
≈ 1,400
≈ 6,000
Qwen2.5 32B
Vendor: Alibaba Cloud
What it does: reads very long documents in one pass — a hundred pages or more — so a whole report can be summarised without being split first. Strong across European and Asian languages.
Requirement
Minimum
Medium
High
GPU type
RTX 4090 (reduced precision)
1× 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)
≈ 420
≈ 1,300
≈ 5,000
Mistral Small 3
Vendor: Mistral AI
What it does: a compact model that summarises at high volume on a single card. The usual choice for summarising incoming email, tickets or news at a steady rate rather than on demand.
Requirement
Minimum
Medium
High
GPU type
RTX 4090
1× L40S 48 GB
1× 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)
≈ 700
≈ 1,900
≈ 5,500
Gemma 2 27B
Vendor: Google
What it does: produces conservative summaries that stay close to the source wording, which reduces the risk of a figure being restated incorrectly. Suited to regulated material.
Requirement
Minimum
Medium
High
GPU type
RTX 4090 (reduced precision)
1× L40S 48 GB
2× H100 80 GB
VRAM
20 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)
≈ 500
≈ 1,500
≈ 5,200
Phi-4 14B
Vendor: Microsoft
What it does: a small model that holds up well on documents dense with numbers, so quarterly figures and measurements survive the summary. Cheap enough for very large archives.
Requirement
Minimum
Medium
High
GPU type
RTX 3090
RTX 4090
1× L40S 48 GB
VRAM
16 GB
24 GB
48 GB
vCPUs
8
12
24
RAM
32 GB
48 GB
96 GB
Server
1× RTX 3090 24 GB
1× RTX 4090 24 GB
1× L40S 48 GB
Rate (pages/hour)
≈ 900
≈ 2,600
≈ 7,000
BART-large-CNN
Vendor: Meta
What it does: an older, purpose-built summarisation model that rewrites news-length passages into short abstracts. It cannot follow instructions, but it is small, fast and predictable, which suits a fixed summarising step inside a larger pipeline.
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)
≈ 2,400
≈ 9,000
≈ 26,000
LED (Longformer Encoder-Decoder)
Vendor: Allen Institute for AI
What it does: a purpose-built model for very long single documents such as court filings and scientific papers, able to read tens of thousands of words at once without splitting. Slower per document than a general model but consistent on long input.
Requirement
Minimum
Medium
High
GPU type
RTX 3090
RTX 4090
1× A100 80 GB
VRAM
16 GB
24 GB
80 GB
vCPUs
8
12
24
RAM
32 GB
48 GB
96 GB
Server
1× RTX 3090 24 GB
1× RTX 4090 24 GB
1× A100 SXM 80 GB
Rate (pages/hour)
≈ 600
≈ 2,000
≈ 5,000
Choosing between them
The decision turns on document length, how much rewriting you will tolerate, and whether summaries are read by staff or fed into another system. Our consultants review your documents and the use the summaries are put to, then recommend the model and the pipeline — single pass or staged — 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 Text summarization. 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.