How Many Languages Does Machine Translation Cover? 464 Across Nine Systems, About 6% of Living Languages

Machine translation providers like to announce their language counts: a hundred languages here, two hundred there, "over 400" in a research release. The numbers are hard to compare, so we did the count ourselves. On 16 November 2026 we downloaded the published language lists of nine systems, normalised every entry to an ISO 639-3 language code and merged the results. The short answer: together they list 464 distinct languages, which is roughly 6% of the languages people speak today.
What we counted
We chose nine systems that publish a machine-readable or clearly tabulated list of languages, covering the three ways machine translation reaches people:
- Commercial cloud APIs: Google Cloud Translation (the language table of its neural machine translation model), Microsoft Translator (its public
/languagesendpoint), the DeepL API and Amazon Translate. - Open-source tools that run on a user's own machine: Firefox Translations (the models Firefox downloads for its built-in, offline page translation, desktop release channel), Argos Translate (the engine behind LibreTranslate) and Apertium, the rule-based platform whose public server lists 133 language pairs.
- Research models with openly released weights: Meta's NLLB-200 ("No Language Left Behind", 2022) and Google's MADLAD-400 translation model (2023).
The raw lists do not count the same thing. DeepL lists 126 codes, but twelve of them are regional variants such as British and American English or Brazilian and European Portuguese. NLLB-200 lists 202 codes, several of which are the same language in two scripts, such as Kashmiri in Arabic and Devanagari script. Apertium lists seven codes for Catalan alone, for spelling norms and the Valencian standard. Merging them lowers every count.
| System | Type | Entries listed | Distinct languages |
|---|---|---|---|
| MADLAD-400 (3B MT model) | Research model | 419 | 416 |
| NLLB-200 | Research model | 202 | 194 |
| Google Cloud Translation | Cloud API | 194 | 187 |
| Microsoft Translator | Cloud API | 138 | 130 |
| DeepL API | Cloud API | 126 | 114 |
| Amazon Translate | Cloud API | 75 | 71 |
| Firefox Translations | Open-source tool | 58 | 57 |
| Argos Translate | Open-source tool | 50 | 48 |
| Apertium (public server) | Open-source tool | 60 | 46 |
The total: 464 languages, about 6% of the world's
Merged, the nine lists contain 464 distinct languages. A few are not living languages, such as Latin, Sanskrit, Esperanto and Klingon. Counting only living languages that have their own individual code in ISO 639-3, the international standard maintained by SIL International, gives 441. The ISO 639-3 code table we downloaded on the same day lists 7,021 living individual languages, so all nine systems together cover about 6.3% of them.
That figure is an upper bound on what people can actually use. If we leave out the two research models and keep only the seven systems that serve translations today, the union falls to 209 languages, 197 of them living individual languages: 2.8% of the ISO total. Google Cloud Translation alone accounts for most of that reach, with 187 languages; the other six deployed systems add only 22 languages that Google's table does not list, among them Upper and Lower Sorbian, Faroese, Inuktitut and Tahitian (Microsoft), Dari (Microsoft, DeepL and Amazon) and Asturian, Arpitan and Northern Sami (Apertium).
A small shared core and a long tail
The overlap between systems is thin. Only 22 languages appear in all nine lists: Standard Arabic, Bulgarian, Catalan, Danish, Dutch, English, French, Hindi, Indonesian, Italian, Norwegian Bokmål, Polish, Portuguese, Romanian, Russian, Slovenian, Spanish, Standard Malay, Swedish, Turkish, Ukrainian and Urdu. German and Japanese miss the list only because Apertium has no pair for them. The four commercial APIs share 66 languages.
| Number of systems listing the language | Languages |
|---|---|
| 9 (all) | 22 |
| 8 | 27 |
| 5 to 7 | 48 |
| 2 to 4 | 113 |
| 1 only | 254 |
At the other end, 254 languages are listed by a single system, and 220 of those by MADLAD-400 alone. NLLB-200 is the only system to list 24 languages, including seven spoken varieties of Arabic (Egyptian, Moroccan, Tunisian, Levantine and others), Kabyle, Santali, Tok Pisin and Central Atlas Tamazight. Google is the only one to list Abkhazian, Ga, Chiga and Hunsrik, a German-based language of southern Brazil. Coverage of widely spoken African languages varies a great deal: Swahili and Hausa appear in five or more systems, but Yoruba, one of the main languages of Nigeria, is absent from DeepL, Amazon and all three open-source tools.
A listed language is not a working language
A language count measures a promise, not a quality. The MADLAD-400 model card is explicit about it: the model was trained on web data covering more than 400 languages, but its parallel training data covered 157 languages and its authors evaluated it on only 204 of the languages it lists. Many of the 220 languages that only MADLAD-400 lists are therefore supported in the sense that the model will produce output, without published evidence of how good that output is. NLLB-200 was evaluated on its full list through the FLORES-200 benchmark, but it is released under a non-commercial licence (CC BY-NC 4.0) and is meant for research.
The same caution applies to commercial systems. Quality for a language depends on how much translated text exists for it, which is the subject of our article on parallel corpora. To check a system for a given language, automatic metrics such as chrF and COMET help, but they need reference translations; our guide to machine translation evaluation explains their limits.
Why the 6% matters
Machine translation is neither possible nor necessarily wanted for every language, but the languages left out are not only small ones. They include regional and minority languages of Europe, most languages of Nigeria, Indonesia and Papua New Guinea, and most Indigenous languages of the Americas: Navajo, for instance, appears only in MADLAD-400's list. The gap follows the availability of digital text, which is why we describe in a separate article what "low-resource" means and which methods work.
Methodology
Sources and dates. All lists were retrieved on 16 November 2026: Google's Cloud Translation language support page (table "Neural Machine Translation model"); the Microsoft Translator endpoint api.cognitive.microsofttranslator.com/languages?api-version=3.0&scope=translation; the DeepL API "Languages supported" page; the Amazon Translate "Supported languages and language codes" page; Mozilla's Remote Settings collection translations-models (records available to the desktop release channel); the Argos Translate package index (argospm-index); Apertium's public server (apertium.org/apy/listPairs); and the Hugging Face model cards of facebook/nllb-200-distilled-600M and google/madlad400-3b-mt. The denominator is the ISO 639-3 code table published by SIL International (file last updated on 15 October 2026): individual languages of type "living".
Normalisation. Every code was mapped to an ISO 639-3 code. Script variants (Serbian in Cyrillic and Latin, Chinese Simplified and Traditional) and regional variants (Canadian French, Brazilian Portuguese) count once. Where a system uses a two-letter code for an ISO macrolanguage, we assigned it to the standard written variety: Arabic to Standard Arabic, Malay to Standard Malay, Persian to Iranian Persian, Swahili to Swahili (individual language), Akan to Twi, and so on. Languages that ISO treats as separate stay separate, even when they are close: Bosnian, Croatian and Serbian; Norwegian Bokmål and Nynorsk; Dari and Persian. Filipino was counted as Tagalog. Spoken Arabic varieties listed by NLLB-200 count as distinct languages, as they have distinct ISO codes. MADLAD-400's generic code for Berber languages is kept as one entry; its placeholder codes "multilingual" and "no linguistic content" were dropped.
Limits. A different counting rule gives different totals: merging all ISO macrolanguages would lower the counts, keeping every Arabic or Quechua variety separate would raise them. Apertium translates between specific pairs, so a language counts here if it appears in at least one pair. Some providers support more languages in other products than in the API we counted. These are declared lists, not quality tests.
Data. The full table, one row per language with a column for each system and the codes as declared, is available as CSV: mt-language-coverage-2026-11.csv (464 rows). Free to reuse with a link to this page; please cite "Parallel Text, machine translation language coverage, November 2026".
Sources
- Google Cloud, "Language support", Cloud Translation documentation, consulted 16 November 2026.
- SIL International, "ISO 639-3 code tables", file updated 15 October 2026.
- NLLB Team and colleagues, "No Language Left Behind: Scaling Human-Centered Machine Translation", arXiv 2207.04672, 2022; Kudugunta and colleagues, "MADLAD-400: A Multilingual and Document-Level Large Audited Dataset", arXiv 2309.04662, 2023; Microsoft Translator, DeepL, Amazon Translate, Mozilla, Argos Translate and Apertium documentation as listed in the methodology.
Multilingual & low-resourceLow-resource languages in NLP
Most of the world's 7,000-plus languages have almost no digital data. What that means for NLP, why web crawls fail them, and what actually helps.
Methods & evaluationEvaluating machine translation
What BLEU really measures, why learned metrics replaced it in many studies, and how professional MQM evaluation catches the errors scores miss.
Multilingual & low-resourceParallel corpora
From the Canadian Hansard and Europarl to EU translation memories and web-mined bitext: the translated text behind machine translation, and its pitfalls.