PaddleOCR languages: every lang code and the model it loads

Find your lang code#

Search by English name, native name, ISO code or script. Each result shows the code PaddleOCR 3.7 accepts, the two models it loads, whether it works in the browser, and a line of Python to copy. Try two languages at once, such as en + fr or Korean + English.

How PaddleOCR turns lang into a model#

In PaddleOCR 3.7 the lang parameter does one thing: it picks a detection model and a recognition model. The logic lives in _get_ocr_model_names() and runs in this order:¹

1. No lang, no ocr_version#

   You get PP-OCRv6_medium_det + PP-OCRv6_medium_rec, which read 50 languages.

2. A PP-OCRv6 language#

   ch, chinese_cht, en, japan and every Latin-script code except pi (52 codes, 50 languages — french and german are aliases of fr and de). All of them load the same two medium models as the default.

3. Another supported script#

   Korean, Thai, Greek, Telugu, Tamil and the Arabic, East Slavic, Cyrillic and Devanagari groups, plus Pali, get PP-OCRv5_server_det and a script-specific *_PP-OCRv5_mobile_rec (59 codes).

4. ka#

   The only code still served by PP-OCRv3: PP-OCRv3_mobile_det + ka_PP-OCRv3_mobile_rec.

5. Anything else#

   PaddleOCR() raises ValueError: No models are available for lang='…' and ocr_version=None.

Three consequences that the official docs don't spell out:

• For PP-OCRv6 languages, lang changes nothing. PaddleOCR(lang="fr"), PaddleOCR(lang="en") and PaddleOCR() load identical models and give identical output.¹⁸
• lang never selects PP-OCRv6 small or tiny. To use them, pass text_detection_model_name and text_recognition_model_name; PaddleOCR then ignores lang with a warning. See selecting a tier.¹
• lang plus ocr_version gives older models. With ocr_version="PP-OCRv5", en loads the English-only en_PP-OCRv5_mobile_rec and fr loads latin_PP-OCRv5_mobile_rec. PP-OCRv4 supports only ch and en; PP-OCRv3 supports 110 codes (all except th and el).¹

Codes that raise ValueError#

PaddleOCR's codes are its own: some match ISO 639-1 (fr, ar, hi), many don't. We ran each of these through PaddleOCR(lang=…) in 3.7.0 and every one failed with ValueError: No models are available for lang=…. The first six come straight from the official PP-OCRv5 language list.³⁴⁸

Language support by script#

Chinese and Japanese#

Use ch for simplified Chinese, chinese_cht for traditional and japan for Japanese — or nothing, since the default PP-OCRv6 model reads all three plus English. On Baidu's in-house benchmark PP-OCRv6 medium scores 90.5% on Japanese recognition versus 73.7% for PP-OCRv5 server.⁵ In our own test, medium and small read every line of our Japanese notice and our Chinese + English notice correctly.⁹

English, French, German, Spanish and other Latin-script languages#

All 46 Latin-script languages plus English are read by the same PP-OCRv6 model: Afrikaans, Albanian, Azerbaijani, Basque, Bosnian, Catalan, Croatian, Czech, Danish, Dutch, Estonian, Finnish, French, Galician, German, Hungarian, Icelandic, Indonesian, Irish, Italian, Kurdish, Latin, Latvian, Lithuanian, Luxembourgish, Malay, Maltese, Maori, Norwegian, Occitan, Polish, Portuguese, Quechua, Romanian, Romansh, Serbian (Latin), Slovak, Slovenian, Spanish, Swahili, Swedish, Tagalog, Turkish, Uzbek, Vietnamese and Welsh.¹² Two quirks:

• Pali (pi) is in the Latin list but explicitly excluded from PP-OCRv6, so it loads latin_PP-OCRv5_mobile_rec instead. Forcing ocr_version="PP-OCRv6" with lang="pi" raises ValueError.¹
• Kurdish (ku) appears in both the Latin and the Arabic-script lists. The Latin route is checked first, so lang="ku" always gets the Latin-script PP-OCRv6 model. For Kurdish in Arabic script, set text_recognition_model_name="arabic_PP-OCRv5_mobile_rec" yourself.¹²

Korean#

lang="korean" loads korean_PP-OCRv5_mobile_rec, which Baidu lists as reading Korean and English, at 88.0% accuracy on a 5,007-image Korean set — a 65.0% improvement over the previous generation, in Baidu's wording.⁴ PP-OCRv6 does not include Korean, and neither does PaddleOCR.js.

from paddleocr import PaddleOCR

ocr = PaddleOCR(lang="korean")   # PP-OCRv5_server_det + korean_PP-OCRv5_mobile_rec
print(ocr.predict("menu_ko.jpg")[0]["rec_texts"])

Arabic, Persian, Urdu and other Arabic-script languages#

Seven codes — ar, fa, ur, ug, ps, sd, bal — load arabic_PP-OCRv5_mobile_rec. Baidu reports 81.27% accuracy on a 2,676-image set (a 22.83% improvement over the previous generation) and lists the model as reading Arabic, Persian, Uyghur, Urdu, Pashto, Kurdish, Sindhi, Balochi and English.⁴ We haven't yet tested Arabic-script documents ourselves, so check the reading order of rec_texts on your own right-to-left pages before relying on it.

Russian, Ukrainian and other Cyrillic languages#

Russian (ru), Ukrainian (uk) and Belarusian (be) load the East Slavic model eslav_PP-OCRv5_mobile_rec (81.6% on 7,031 images). The other 30 Cyrillic codes — Bulgarian, Serbian (rs_cyrillic), Macedonian, Kazakh, Kyrgyz, Mongolian, Tatar, Chechen and more — load cyrillic_PP-OCRv5_mobile_rec (80.27% on 7,600 images).¹⁴

Hindi, Marathi, Nepali and other Devanagari languages#

Thirteen codes load devanagari_PP-OCRv5_mobile_rec: hi (Hindi), mr (Marathi), ne (Nepali), sa (Sanskrit), bh, bho, mai, mah, new, gom, bgc, ang and sck. Baidu reports 84.96% accuracy on 3,611 images, a 68.26% improvement over the previous generation.⁴ The docs disagree on two names: ang is "Old English" in the language tables but "Angika" in the model's language list, and sck is "Sadri" in one and "Santali" in the other.³⁴

from paddleocr import PaddleOCR

ocr = PaddleOCR(lang="hi")   # PP-OCRv5_server_det + devanagari_PP-OCRv5_mobile_rec

Thai, Greek, Tamil, Telugu#

Each has its own PP-OCRv5 recognizer that also reads English: th (82.68%), el (89.28%), ta (94.2%) and te (87.65%), on Baidu's per-language test sets of 2,121 to 4,261 images.⁴

The ka problem: Georgian or Kannada?#

The OCR pipeline docs list ka as Georgian. But the only model it loads, ka_PP-OCRv3_mobile_rec, is described in two other places in the same docs as a Kannada model — a different language with a different script.³⁶ There is no PP-OCRv5 or v6 model for either. We couldn't resolve this from the sources, so test lang="ka" on a sample of your documents before using it for Georgian or Kannada.

Speed and size for languages outside PP-OCRv6#

Every PP-OCRv5 language route pairs a small script recognizer with the large PP-OCRv5_server_det detector (84.3 MB in the docs). The recognizers themselves are small — Baidu describes the October 2025 batch as having "only 2M parameters" — so on CPU the detector does most of the work.⁴¹⁰

That matters if you use ONNX Runtime. In our benchmark, the PP-OCRv5 server pipeline averaged 35.7 s per page under ONNX Runtime against 5.2 s with PaddlePaddle 3.2.0. We re-checked with lang="korean" on our receipt image (Korean model, English text, which it is listed as reading): 12.35 s per image with ONNX Runtime vs 1.80 s with PaddlePaddle, median of three warm runs, same 27 lines of output.¹¹ Because the Korean recognizer is a small mobile model, this suggests the slowdown comes from the server detector under ONNX Runtime — the opposite of PP-OCRv6, where ONNX Runtime was the faster engine in every tier.

Mixed-language documents#

One PaddleOCR instance runs one recognition model. lang takes a single code; strings like "en+fr" or "en,fr" raise ValueError.⁸ What to do depends on the mix:

For the last case, one approach is to run both pipelines on the same image and, for each text region, keep the line with the higher rec_scores value. This is our suggestion, not something we've benchmarked; a recognizer reading a script it doesn't know can still return a confident wrong answer, as PP-OCRv6 tiny on Japanese shows.

from paddleocr import PaddleOCR

common = dict(use_doc_orientation_classify=False, use_doc_unwarping=False,
              use_textline_orientation=False)
ko = PaddleOCR(lang="korean", **common)   # korean_PP-OCRv5_mobile_rec
ru = PaddleOCR(lang="ru", **common)       # eslav_PP-OCRv5_mobile_rec

a = ko.predict("sign.jpg")[0]
b = ru.predict("sign.jpg")[0]
for name, r in (("korean", a), ("russian", b)):
    for text, score, box in zip(r["rec_texts"], r["rec_scores"], r["rec_boxes"]):
        print(name, round(score, 3), box.tolist(), text)

If you have a GPU and many mixed-script pages, PaddleOCR-VL is the other route: Baidu says it supports 109 languages in a single vision-language model.¹⁰

Which languages work in the browser#

The official browser SDK, PaddleOCR.js, ships three model sets: PP-OCRv6 small, PP-OCRv6 tiny and PP-OCRv5 mobile.⁷

• ocrVersion: "PP-OCRv6" loads PP-OCRv6 small and accepts ch, chinese_cht, en, japan and the Latin-script codes except pi — the same 50 languages as Python.
• ocrVersion: "PP-OCRv5" accepts only ch, chinese_cht, en and japan. Any other code throws Unsupported lang/ocrVersion combination.
• Tiny has no lang mapping; you load it by model name, and it can't read Japanese.

Korean, Arabic, Cyrillic, Devanagari, Thai, Greek, Tamil and Telugu are Python-only for now. Our online OCR tool uses PaddleOCR.js, so it reads exactly the PP-OCRv6 set; the finder above marks each language accordingly. Embedding details are in the PaddleOCR.js guide.

Errors in the official language lists#

We compared the documentation against langs.py and ocr.py. The code is what runs, so the finder and table on this page follow the code.

• The PP-OCRv5 list in the OCR pipeline docs contains rslatin, sr, ab, av, ce and lki, none of which the code accepts, and omits the accepted french, german, abq, ava, che, lbe, rs_latin and rs_cyrillic.³
• ab is called Abkhaz in the PP-OCRv5 docs, while the accepted abq is called Abaza in the OCR docs — two different languages.³⁴
• ka is Georgian in one table and Kannada in the model description (above).
• The "Supported Languages" table in the OCR docs has 81 rows; the code accepts 112 codes. Missing from the table: Thai, Greek, Catalan, Basque, Finnish, Galician, Luxembourgish, Romansh, Quechua, Pashto, Sindhi, Balochi and 18 Cyrillic-script languages including Kazakh, Kyrgyz, Tajik, Macedonian and Tatar.³
• The documented PP-OCRv3 list omits 27 codes that the code does route to PP-OCRv3 models — for example kk → cyrillic_PP-OCRv3_mobile_rec.¹³
• Language counts vary: PP-OCRv5 multilingual is "106 languages" in its own doc and "109 languages" in the README.⁴¹⁰

All 112 lang codes#

Sorted by language name. "Recognition model" is what PaddleOCR 3.7.0 loads when you pass only lang; the detector is PP-OCRv6_medium_det for PP-OCRv6 rows, PP-OCRv5_server_det for PP-OCRv5 rows and PP-OCRv3_mobile_det for ka. "v6 tiny" says whether PP-OCRv6 tiny reads the language; "Browser" whether PaddleOCR.js and our online tool do. Generated from the 3.7.0 source; the same data drives the finder.¹²

Questions#