Can AI summarise a 100-page PDF? Yes, if it can read the PDF, and if you check what it leaves out.
Short answer. Yes, with two conditions. First, the PDF must have a text layer, or be scanned and read by OCR first. Second, the model must be able to hold the text. A 100-page report at 400 to 500 words a page is roughly 53,000 to 67,000 tokens in English. That fits the 128K-token windows of Gemma 3 27B and Llama 3.3 70B, but is more than the 32,768 tokens Qwen3-32B takes natively. Fitting is not the same as reading well: research on long inputs finds that models miss details in the middle and that many cannot use their full claimed window. Treat the summary as a map and check it against the pages.
01
Will 100 pages fit in the model?
A model reads in tokens, which are pieces of words. OpenAI’s documentation gives a rough rule for English: one token is about 4 characters or 0.75 words. At 400 to 500 words a page, 100 pages come to 40,000 to 50,000 words, or roughly 53,000 to 67,000 tokens, before the model writes anything back.
Hindi and other Indian-language text can take many more tokens per word. One study found that the same text translated into different languages can need up to 15 times as many tokens with some tokenizers, so a 100-page report in an Indian language may not fit where the English one does. The context window also costs memory, which the KV cache entry in the glossary explains.
Open model
Context window (model card)
Holds about 60,000 tokens of English?
Qwen3-32B
32,768 tokens natively, 131,072 with YaRN scaling
Not natively. Yes with YaRN switched on.
Gemma 3 27B
128K tokens in, 8,192 tokens out
Yes
Llama 3.3 70B
128k tokens
Yes
02
Where long-input summaries go wrong
Lost in the Middle (Liu and others) found that models often do best when the relevant fact sits at the start or end of the input, and do markedly worse when it sits in the middle, even models built for long inputs. RULER (Hsieh and others) tested 17 models that all claimed windows of 32K tokens or more; only about half kept satisfactory performance at 32K.
FABLES (Kim and others) checked summaries of 26 recent books of over 100,000 tokens. Annotators who had read each book judged 3,158 claims. The unfaithful claims were mostly about events and the state of characters, and they usually took indirect reasoning to spot. LLM-based checkers did not agree well with the human judges. In an office report the equivalent is a claim that depends on two clauses read together, such as an exception on page 61 to a rule on page 12. A summary can state the rule and miss the exception.
03
What a person must still check
Numbers, units and dates. Compare each with the page it came from.
Negatives and exceptions: not, unless, except, subject to.
Defined terms. A word defined on page 3 may mean something narrower than its everyday sense.
Tables and footnotes, which are the parts most often skipped or flattened.
What the summary leaves out. Ask the model to list the sections it covered least, then read those yourself.
Anything the summary will be used to decide. Read the source passage.
04
How an office would set this up
Check that the PDF has text: try selecting a line in a PDF viewer. If you cannot, the pages are images and need OCR first, as the Camelot documentation also advises.
Extract the text page by page and keep the page numbers.
Split the text at its headings into sections of 5 to 15 pages and summarise each section with page references.
Merge the section summaries into a one-page summary, then ask separately what the summary leaves out.
Spot-check five claims by finding them in the PDF text.
Store the summary beside the PDF with the model name and the date.
05
What ANVE does today
ANVE’s documents function answers questions over the office’s own documents and shows the source passage for each answer. Its vision models read scanned documents with OCR. ANVE does not list a dedicated long-document summariser, and full document drafting with design templates is not on ANVE yet.
The documents function runs today in the development build. No sealed unit exists yet.
06
Questions people ask
How many tokens is a 100-page PDF?
Roughly 53,000 to 67,000 tokens in English at 400 to 500 words a page, using the rule of thumb of 0.75 words per token. Text in Indian languages can need many more.
Can it summarise a scanned PDF?
Only after the pages have been read by OCR, or by a vision model that can read images of pages. Check the OCR output on a few pages before trusting the summary.
Is it better to ask questions than to ask for a summary?
For checking specifics, often yes. A question answered from retrieved passages with the source shown is easier to verify than a summary of everything.
Which open models handle long documents best?
It depends on the window, the memory you have and the task. The separate question on long-document models compares them, and your own test on your own documents decides.
07
What is not real yet
No sealed ANVE unit exists; every image of the appliance on this site is a labelled render. The software runs today on a development machine. No customers yet, no certifications, no delivery date, and the final price is not announced.
Reservations are ₹999 in India or US$12 elsewhere, refundable in full until you confirm an order in writing (terms). Offices that cannot use cloud AI can apply to be one of ten design partners.
