Every Spanish, French, Italian, Portuguese word list

I needed frequency lists for four languages, and I spent longer on the licences than on the data.

That's the wrong ratio, and the reason for this page. Every roundup says where to download these lists.

Almost none says what you may legally do with them. Several of the most-recommended ones forbid commercial use in a file you have to go looking for.

So this page carries a licence column. I read every cell at its own source on 3 September 2026, instead of copying it from another table. Where I couldn't reach a source, the cell says so.

Several popular lists forbid commercial use, in a file you have to find

Start with the ones most likely to end up inside a language app, because those are the ones that will bite you.

The SUBTLEX family covers all four: ESP for Spanish, IT for Italian, PT and PT-BR for Portuguese. French has none, because Lexique already did that job.

All sit on OSF (the Open Science Framework, a site where researchers post data). Every OSF licence field I checked says only "Other" and points at a file. I downloaded the files.

SUBTLEX-ESP's reads:

NonCommercial. If you sell a language app, that licence doesn't cover you. The OSF project page prints those four words nowhere a browser will show you: they live in a 471-byte text file you have to download.

SUBTLEX-IT says the same in 109 bytes. SUBTLEX-PT-BR goes further than any other list here:

NoDerivatives. On a word list, that plausibly rules out the ordinary thing you'd do with one: reformat it, cut it to the top 5,000, add a translation column.

I say plausibly because I haven't seen anyone test what counts as a derivative of a frequency table, and the answer isn't settled.

SUBTLEX-PT has no licence file and no OSF licence set. Silence isn't permission, and a missing licence is the cell I'd email an author about.

CEFRLex does the same more visibly. Its Spanish ELELex and French FLELex download pages both state CC BY-NC-SA 4.0 in the page body. Same restriction, better signposting.

Now the one the internet gets backwards. The hermitdave OpenSubtitles lists are the most-recommended free option for all four languages, and they're widely described as MIT. The repository's own README splits it:

MIT covers the generator. The word lists themselves are share-alike, which is a real obligation: what you make from them has to go out under the same licence.

The repository's LICENSE file contains only the MIT text, which is how the misreading starts. People open LICENSE, see MIT, and never reach the README.

One more thing about that repository, since most people use it. Its content directory holds two vintages, 2016 and 2018, so the prebuilt lists are eight years behind their own corpus.

For learner vocabulary that hardly matters, because the top 5,000 words don't churn. For anything sensitive to current usage, it matters a lot.

I'm not a lawyer, and this isn't legal advice. But the gap between "do what you like with attribution" and "you may not sell this" is worth four minutes with the file.

Every list side by side, and the unit column decides which you can use

The unit column is the one that decides whether a list is usable, and the next section is why.

One thing is missing on purpose. The Leipzig Corpora Collection belongs there, and I couldn't get in.

Every entry point I tried on 3 September 2026 returned an Anubis proof-of-work challenge (a bot check), and the direct download host returned 403. I won't print a licence I couldn't read, so Leipzig has no row.

If you can reach it in a browser, it covers all four languages and is worth the trip.

The RAE row's blank licence cell is the same story: RAE publishes CREA frequency lists and, as far as I can find, no terms of use for them.

Form lists scatter one Spanish verb across the whole ranking

Most of these lists count surface forms (each spelling as it appears in text), which the download pages rarely say plainly. Every inflection of a verb is a separate row competing for its own rank. Spanish conjugates heavily, so this gets expensive fast.

I took the Spanish 50k list and looked up 48 forms of hablar. 38 of them are in there. Here's where they sit:

One Spanish verb, scattered across a frequency list

0

separate rows the verb hablar occupies

0

rank of its most common form

0

rank of the lowest form it still owns

Treated as one word, hablar is the 57th most common thing in the corpus. Treated as forms, no row it owns cracks the top 150.

A learner working down the list collects hablas at 774 and hablo at 1,045, and is still picking up stragglers past 47,000. One verb, smeared across the whole range and quietly demoted.

French does the same with parler: 25 forms present, best rank 143, worst 45,368. Every form list for a Romance language has this property. It's what counting forms means.

Which is why Lexique is the standout here. It ships the form and its lemma as columns in one downloadable file. (A lemma is the dictionary form, the way hablar is the lemma of hablas.)

It also splits film frequency from book frequency, so you can see which register a word belongs to. I downloaded both: Lexique383.tsv is 25.8 MB and 142,695 rows, Lexique400.tsv 33 MB and 189,863.

Spanish has a partial answer in EsPal. It returns lemma and part-of-speech properties for words you paste in, over a written corpus or over opensubtitles.org tokens.

It solves the unit problem as a lookup service rather than a file: no bulk download, databases dated 2012.

The other route, in any language here, is to lemmatise a form list yourself. spaCy publishes news models for all four (es_core_news_sm and its siblings) under MIT. simplemma is MIT too, and faster if you only want lemmas.

Collapsing a 50k list is a short script. What it costs you is responsibility for the tagger's mistakes (the tagger is the software that labels each word's grammar).

A tagger that guesses wrong on porto, which is both a noun and a form of portare, quietly corrupts the rank of two different words at once.

The word-family question sits one level above this, and it changes the numbers again. The unit the famous coverage studies actually count is neither forms nor lemmas.

The top 1,000 words cover more of French than of Spanish

Since I had all the full lists open, I summed them. I can't find these published anywhere, so check my arithmetic before leaning on it.

French runs about 4.5 points ahead of Spanish at 1,000 words, with the same corpus and the same method. That's a real difference in how top-heavy the two languages are.

And Brazilian Portuguese has its own list with roughly twice the tokens (running words) of the European one. So pt_br is the file you want if that's your variety.

Each percentage is coverage of the corpus the list was built from. Subtitles are dialogue, so these describe speech rather than print, and the shape of that curve matters more to a learner than the exact figure.

Two of these lists have stopped updating, and both say so

The two biggest general-purpose options both have a date past which they won't move, and both say so in public.

wordfreq is frozen at 2021, deliberately. Its README says the data "is unlikely to be updated again," and the maintainer wrote a document explaining why. The first reason:

Generative AI text flooded the open web wordfreq sampled. Her position is that updating would make the data worse, so she stopped.

The same document refuses CSV export outright, because a CSV has nowhere to carry the share-alike attribution.

CREA is frozen harder and more quietly. RAE's own page gives the numbers:

Version 3.2, June 2008, 160 million forms, texts stopping in 2004. So the RAE Spanish frequency list circulating today samples a language that ends 22 years ago, before smartphones.

RAE's live successor CORPES XXI passed 395 million forms at version 1.0 in May 2023. It publishes no downloadable frequency list at all. The downloadable one is the dead one.

RAE's own site served me nothing. rae.es and corpus.rae.es both returned 403 to every request on 3 September 2026, so the quotes above come from Internet Archive captures.

The frequency-list page they preserve offers four files: 1,000, 5,000 and 10,000 forms, plus a full list as a zip. Most descriptions mention only the single 1,000-word list.

Which list to take depends on what you're building

Building something you'll sell: hermitdave, and honour the share-alike. It covers all four with a licence that permits commercial use. wordfreq's CC BY-SA data does too, and it blends several sources instead of only subtitles, but it's frozen at 2021 and ships as a Python package rather than plain files.

Skip the SUBTLEX lists and CEFRLex whatever their quality, because the licence answers the question before the data does.

Working in French: Lexique, and it isn't close. Lemmas and forms in one file, film and book frequencies separated, CC BY-SA. Working in Spanish and willing to paste rather than download: EsPal.

Teaching, or building a course with levels: CEFRLex for Spanish and French, KELLY for Italian. Both are CEFR-graded, meaning sorted by the European levels from A1 to C2, and both are noncommercial.

CEFRLex was built from textbooks and graded readers, where the rest use a general corpus. That makes it a different object from the rest of this table, which is worth knowing before you compare its numbers to anyone else's.

Working in Italian: CoLFIS, the least-known list here and the only free lemmatised Italian one here that doesn't forbid commercial use. It ships a lemmario and a formario (a lemma list and a form list) separately, the split that makes Lexique good. It also counts multiword expressions as units.

The catch is the corpus: newspapers, periodicals and books from 1992 to 1994, weighted by ISTAT data on what Italians read. Print rather than speech, and 30 years old.

Its download page says "L'accesso è gratuito" and names no licence, a gap worth an email before you ship it.

The honest hole is Portuguese. There's a Routledge volume, but it's a print book. SUBTLEX-PT states no licence at all, and SUBTLEX-PT-BR forbids derivatives.

The free, clearly-licensed, lemmatised Portuguese list doesn't exist, and I'd rather say so than pad the row.

I hit that wall building Verbamor's four starter decks. They're 646 Spanish rows, 628 French, 623 Italian and 642 Portuguese.

Those numbers differ from one concept list because of the unit problem. pequeño / pequeña is one row for us and two entries in a form list, and the four languages disagree about how often that happens.

Verbamor is a $19.99/month iOS app that turns the words you've already met into scheduled recall. A frequency list is where a beginner starts. The app ships an Essentials deck of about 650 common words, then picks up from your own lessons. If the word lists are all you wanted, a premade deck built from any source above does it for free.

---

Sources