A natural question for any library: how much is in here? We have 6.5 million pages across roughly 15,500 readable books in 158 languages. But pages and books are containers. We wanted the contents — the actual words.
So we counted. Not just the original texts, but everything the pipeline adds: the facing-page English translations, the per-page summaries, the descriptions our vision models write for 142,000 illustrations. The answer is about seven billion words — or, if you only trust the most conservative count, about five. Either way, that puts Source Library at roughly the scale of English Wikipedia.
Against the obvious yardstick
Wikipedia is the reference everyone has a feel for. English Wikipedia is about 4.7 billion words of article prose; all 300-plus language editions together come to roughly 30 billion. Here is where we land, counting our original texts and their translations as two layers of the same bar:
Original-language OCR + AI English translation + enrichment
Counting originals and translations together, we are a little larger than English Wikipedia. Counting only the original source texts — the centuries-old Latin, Greek, Chinese, Sanskrit, Arabic, and the rest — we are about two-thirds of it. And against the whole multilingual Wikipedia, we are something like a quarter. For a library of pre-modern primary sources, assembled in a few years, that is a strange and slightly vertiginous fact.
What's actually in the count
The number is the sum of distinct layers. Roughly half is the original text, scanned and OCR'd page by page. Almost as much again is the AI translation that sits beside it — a little larger per page, because Latin and Greek are compact and English spells everything out. The enrichment layer (summaries, keywords, the museum-style descriptions for every illustration) is real but small.
~6.2M pages · 158 languages
~5.5M pages translated
summaries + 142K image descriptions
Why we give a range, not a number
The honest answer is a range — five to seven billion words — and the gap between those two figures is itself a story about machine translation.
When we first averaged the words per translated page, the number came out absurdly high: nearly 1,900 words a page, which would have implied a translation corpus of over ten billion words on its own. A typical page has about 460. Something was dragging the average up by a factor of four.
The culprit was a familiar failure mode. About 2.4% of pages — on the order of 150,000 of them — carry translations that run away. The model, an older preview build, hits a repetitive passage and gets stuck, emitting the same phrase tens of thousands of times until it slams into its output ceiling. One leaf of a Tibetan sutra translated into the sentence “They see the Dharma.” repeated 6,522 times — 52,000 words from a page whose original is 1,600. (We wrote about the most extreme case, a Javanese Bible leaf, in Does the AI Get Religion?)
A handful of 50,000-word junk pages can fabricate billions of words that no reader would ever see. So we don't use the raw average. We cap each page's contribution at a generous 3,000 words — more than any real folio — which gives the seven-billion headline. The five-billion floor is the page median, which ignores the long tail entirely. The truth sits between, and both comfortably clear “about the size of Wikipedia.” Those runaway pages are now on the list to be re-translated and purged from search.
The whole measurement is reproducible: scripts/analytics/corpus-size.mjs samples pages from the database, strips the editorial annotation that wraps each one, and reports the median, winsorized, and raw figures side by side — plus a --outliers mode that lists the runaway pages.
Size is not the point of a library — a single readable page of Ficino that no one could read before is worth more than a million words of boilerplate. But scale does say something. A few years ago this was one untranslated book in a glass case in Amsterdam. It is now, by word count, a Wikipedia's worth of primary sources that had mostly never been carried into English — sitting in one place, readable and quotable, page by facing page.
Counts measured 1 June 2026 from the production database. Wikipedia baselines: English ~4.7B words of article prose; all editions ~30B. Word totals are $sample extrapolations (±a few percent); token estimates assume ~1.4 tokens per word for mixed scripts.