Models
Models · Translator · Public release · 16 November 2026

Paar

Twelve small translators from English, 17.7 MB each, that run offline on a CPU or in your browser.

On everyday conversational sentences, English to Gujarati and English to Bengali are ahead of Firefox’s similar-size offline models (+2.11 and +1.32 chrF++), and Telugu by +0.73; Gujarati also on news-style text (+0.66). Paired 95% intervals, holding after correction for all 18 Firefox comparisons.

Paar is twelve small translation models by Udgam Labs, one for each direction: English to Hindi, Bengali, Marathi, Telugu, Tamil, Gujarati, Urdu, Kannada, Malayalam, Punjabi, Odia and Assamese. Each has 17.3M parameters; the int8 model file is 17.7 MB, and it translates a sentence in 29 to 37 ms on one thread of a Mac mini CPU, with no server and no internet connection. The same models run in a web page with transformers.js (tested in Chrome and Safari on a Mac), a 28.7 MB download per language, then offline.

They are distilled from IndicTrans2: each learned from English sentences that IndicTrans2 dist-200M translated. Bigger models translate better. indicMT, about 9 times the size, is ahead on every set we scored except one test set each for Kannada, Malayalam and Tamil, by up to 4.8 chrF++. Firefox’s offline models are ahead in eight of the twelve languages; the table below shows each one.

Try it now · in your browser · free, no login

Type English. Read it in Hindi, Bengali, Marathi, Telugu, Tamil, Gujarati, Urdu, Kannada, Malayalam, Punjabi, Odia or Assamese.

Runs in your browser: about 29 MB once per language, kept for next time. Nothing you type is sent anywhere.

Tested in Chrome and Safari on a Mac.

Hindi

The translation appears here.

The result

English to Gujarati in 17.7 MB, ahead of Firefox’s similar-size offline model on conversational sentences (+2.11) and on news-style text (+0.66).

On FLORES+ devtest it scores 53.21 chrF++ against 52.55 for Firefox’s offline English-to-Gujarati model (17.1 MB): +0.66, paired 95% interval +0.13 to +1.20. On conversational sentences (IN22-Conv) it scores 49.60 against 47.49: +2.11, interval +1.59 to +2.67. Both hold on the int8 build you download (+0.69 and +2.02) and after a Holm correction over all 18 Firefox comparisons; the FLORES+ lead is the smaller of the two. Firefox’s file is 17.1 MB; our int8 download is 19.7 MB. It is 0.74 (FLORES+) and 0.52 (IN22-Conv) behind indicMT. Same sentences and scorer for every system; measured 14 and 16 November 2026.

  • Bengali and Telugu: ahead of Firefox’s similar-size offline models on everyday conversational sentences (IN22-Conv: +1.32 and +0.73, the Telugu lead the smaller), behind them on news-style text (FLORES+: 0.93 and 1.12).
  • indicMT: on FLORES+, Kannada, Malayalam and Tamil show no measurable difference from indicMT, a model nine times their size; it is ahead of them on the other nine FLORES+ sets and on all twelve conversational sets.
  • Punjabi, Assamese and Odia: Firefox has no model for these three; every other open model we found for them is larger.
  • All twelve: about 8.5 times smaller to download than indicMT (19.4 to 20.0 MB against 169.2 MB at int8), and level with it to 4.8 chrF++ behind it, by language and set.
  • Behind Firefox: Hindi, Marathi, Tamil and Urdu on both sets (by 1.0 to 3.6), Kannada on both (1.21 and 0.34) and Malayalam on FLORES+ (0.74). Firefox’s Marathi, Tamil and Urdu models are 31.6 MB.
Size
17.7 MB
The int8 model file, each language; 19.4–20.0 MB with its vocabulary, about 8.5 times smaller to download than indicMT (169.2 MB).
Parameters
17.3M
Each of the twelve. indicMT, the usual open model for these directions, has 153.8M.
Speed
29–37 ms
A sentence, median, on one thread of a Mac mini CPU (CTranslate2 int8).
Ahead of Firefox
+2.11 / +1.32
chrF++ on conversational sentences, Gujarati / Bengali, against Firefox’s similar-size offline models; holds after correction for 18 comparisons.

What it writes

Three everyday sentences in all twelve, from each card’s quick start. These are unedited outputs, mistakes included, and no native speaker has reviewed them. In Gujarati, “by 6:15” became દ્વારા (“by means of”) and 6:45 became 6.45; the battery “dies” as a person does (મૃત્યુ પામે છે).

English → Hindi

  • The train leaves at 6:45 tomorrow morning, so please be at the station by 6:15.

    ट्रेन कल सुबह 6.45 बजे रवाना होती है, इसलिए कृपया स्टेशन पर 6.15 बजे तक रहें।

  • Can you send me the invoice for the last three months?

    क्या आप मुझे पिछले तीन महीनों से चालान भेज सकते हैं?

  • My phone battery dies by the afternoon even though I charge it overnight.

    दोपहर तक मेरी फोन की बैटरी की मृत्यु हो जाती है, भले ही मैं इसे रातोंरात चार्ज करता हूं।

English → Bengali

  • The train leaves at 6:45 tomorrow morning, so please be at the station by 6:15.

    ট্রেনটি আগামীকাল সকাল 6টা 45 মিনিটে চলে যায়, তাই দয়া করে 6টা 15 মিনিটের মধ্যে স্টেশনে থাকে।

  • Can you send me the invoice for the last three months?

    আপনি কি আমাকে গত তিন মাস ধরে চালানটি পাঠাতে পারেন?

  • My phone battery dies by the afternoon even though I charge it overnight.

    আমি রাতারাতি চার্জ করার পরেও বিকেলের মধ্যে আমার ফোন ব্যাটারি মারা যায়।

English → Marathi

  • The train leaves at 6:45 tomorrow morning, so please be at the station by 6:15.

    ट्रेन उद्या सकाळी 6:45 वाजता निघून जाते, म्हणून कृपया 6:15 द्वारे स्थानकावर रहा.

  • Can you send me the invoice for the last three months?

    तुम्ही मला गेल्या तीन महिन्यांपासून चलन पाठवू शकता का?

  • My phone battery dies by the afternoon even though I charge it overnight.

    मी एका रात्रीत चार्ज करत असतानाही माझी फोन बॅटरी दुपारपर्यंत मरते.

English → Telugu

  • The train leaves at 6:45 tomorrow morning, so please be at the station by 6:15.

    రైలు రేపు ఉదయం 6.45 కి బయలుదేరుతుంది, కాబట్టి దయచేసి స్టేషన్లో 6.15 వద్ద ఉంటుంది.

  • Can you send me the invoice for the last three months?

    మీరు గత మూడు నెలలుగా నన్ను ఇన్వాయిస్ పంపగలరా?

  • My phone battery dies by the afternoon even though I charge it overnight.

    నేను రాత్రిపూట ఛార్జ్ చేసినప్పటికీ నా ఫోన్ బ్యాటరీ మధ్యాహ్నం నాటికి మరణిస్తుంది.

English → Tamil

  • The train leaves at 6:45 tomorrow morning, so please be at the station by 6:15.

    ரயில் நாளை காலை 6.45 மணிக்கு செல்கிறது, எனவே தயவுசெய்து 6:15 க்குள் நிலையத்தில் இருங்கள்.

  • Can you send me the invoice for the last three months?

    கடந்த மூன்று மாதங்களுக்கு நீங்கள் எனக்கு விலைப்பட்டியல் அனுப்ப முடியுமா?

  • My phone battery dies by the afternoon even though I charge it overnight.

    நான் ஒரே இரவில் சார்ஜ் செய்தாலும் எனது தொலைபேசி பேட்டரி பிற்பகல்க்குள் இறந்துவிடுகிறது.

English → Gujarati

  • The train leaves at 6:45 tomorrow morning, so please be at the station by 6:15.

    ટ્રેન આવતીકાલે સવારે 6.45 પર રવાના થાય છે, તેથી મહેરબાની કરીને 6:15 દ્વારા સ્ટેશન પર રહો.

  • Can you send me the invoice for the last three months?

    શું તમે મને છેલ્લા ત્રણ મહિના માટે ઇન્વૉઇઝન મોકલી શકો છો?

  • My phone battery dies by the afternoon even though I charge it overnight.

    હું તેને રાતોરાત ચાર્જ કરું છું તેમ છતાં મારી ફોનની બેટરી બપોર સુધીમાં મૃત્યુ પામે છે.

English → Urdu

  • The train leaves at 6:45 tomorrow morning, so please be at the station by 6:15.

    ٹرین کل صبح 6 بج کر 45 منٹ پر چلی جاتی ہے، اس لیے براہ کرم اسٹیشن پر 6:15 تک رہیں۔

  • Can you send me the invoice for the last three months?

    کیا آپ مجھے پچھلے تین مہینوں سے انوائس بھیج سکتے ہیں؟

  • My phone battery dies by the afternoon even though I charge it overnight.

    میری فون بیٹری دوپہر تک مر جاتی ہے حالانکہ میں اسے راتوں رات چارج کرتا ہوں۔

English → Kannada

  • The train leaves at 6:45 tomorrow morning, so please be at the station by 6:15.

    ರೈಲು ನಾಳೆ ಬೆಳಿಗ್ಗೆ 6:45 ಕ್ಕೆ ಹೊರಟು ಹೋಗುತ್ತದೆ, ಆದ್ದರಿಂದ ದಯವಿಟ್ಟು 6:15 ರ ವೇಳೆಗೆ ನಿಲ್ದಾಣದಲ್ಲಿರಿರಿ.

  • Can you send me the invoice for the last three months?

    ನೀವು ಕಳೆದ ಮೂರು ತಿಂಗಳುಗಳಿಂದ ನನಗೆ ಇನ್ವಾಯ್ಸ್ ಕಳುಹಿಸಬಹುದೇ?

  • My phone battery dies by the afternoon even though I charge it overnight.

    ನಾನು ರಾತ್ರಿಯಿಡೀ ಚಾರ್ಜ್ ಮಾಡಿದರೂ ನನ್ನ ಫೋನ್ ಬ್ಯಾಟರಿ ಮಧ್ಯಾಹ್ನದೊಳಗೆ ಸಾಯುತ್ತದೆ.

English → Malayalam

  • The train leaves at 6:45 tomorrow morning, so please be at the station by 6:15.

    ട്രെയിൻ നാളെ രാവിലെ 6.45 ന് പോകുന്നു, അതിനാൽ ദയവായി 6:15 ഓടെ സ്റ്റേഷനിൽ ഉണ്ടാകുക.

  • Can you send me the invoice for the last three months?

    കഴിഞ്ഞ മൂന്ന് മാസമായി നിങ്ങൾക്ക് എനിക്ക് ഇൻവോയ്സ് അയയ്ക്കാൻ കഴിയുമോ?

  • My phone battery dies by the afternoon even though I charge it overnight.

    ഞാൻ ഒറ്റരാത്രികൊണ്ട് ചാർജ് ചെയ്തെങ്കിലും എന്റെ ഫോൺ ബാറ്ററി ഉച്ചകഴിഞ്ഞ് മരിക്കുന്നു.

English → Punjabi

  • The train leaves at 6:45 tomorrow morning, so please be at the station by 6:15.

    ਰੇਲਗੱਡੀ ਕੱਲ੍ਹ ਸਵੇਰੇ 6.45 ਵਜੇ ਰਵਾਨਾ ਹੁੰਦੀ ਹੈ, ਇਸ ਲਈ ਕਿਰਪਾ ਕਰਕੇ ਸਟੇਸ਼ਨ 'ਤੇ 6:15 ਤੱਕ ਰਹੋ।

  • Can you send me the invoice for the last three months?

    ਕੀ ਤੁਸੀਂ ਮੈਨੂੰ ਪਿਛਲੇ ਤਿੰਨ ਮਹੀਨਿਆਂ ਲਈ ਚਲਾਨ ਭੇਜ ਸਕਦੇ ਹੋ?

  • My phone battery dies by the afternoon even though I charge it overnight.

    ਮੇਰੀ ਫੋਨ ਬੈਟਰੀ ਦੁਪਹਿਰ ਤੱਕ ਮਰ ਜਾਂਦੀ ਹੈ ਭਾਵੇਂ ਕਿ ਮੈਂ ਇਸ ਨੂੰ ਰਾਤ ਭਰ ਚਾਰਜ ਕਰਦਾ ਹਾਂ।

English → Odia

  • The train leaves at 6:45 tomorrow morning, so please be at the station by 6:15.

    ଟ୍ରେନ୍ ଆସନ୍ତାକାଲି ସକାଳେ 6:45 ରେ ପ୍ରସ୍ଥାନ କରେ, ତେଣୁ ଦଯ଼ାକରି ଷ୍ଟେସନରେ 6:15 ସୁଦ୍ଧା ରୁହନ୍ତୁ |

  • Can you send me the invoice for the last three months?

    ଆପଣ ମୋତେ ଗତ ତିନି ମାସ ଧରି ଇନଭଏସ୍ ପଠାଇପାରିବେ କି?

  • My phone battery dies by the afternoon even though I charge it overnight.

    ମୁଁ ରାତାରାତି ଚାର୍ଜ କରିବା ସତ୍ତ୍ୱେ ଅପରାହ୍ନରେ ମୋ ଫୋନ୍ ବ୍ଯ଼ାଟେରୀ ମରିଯାଏ |

English → Assamese

  • The train leaves at 6:45 tomorrow morning, so please be at the station by 6:15.

    ৰে "লখন কাইলৈ ৰাতিপুৱা 6ঃ45 ত গুচি যায়, সেয়েহে অনুগ্ৰহ কৰি 6:15 ৰ দ্বাৰা ষ্টেচনত থাকিব।

  • Can you send me the invoice for the last three months?

    আপুনি যোৱা তিনি মাহ ধৰি মোক চালান প্ৰেৰণ কৰিব পাৰিবনে?

  • My phone battery dies by the afternoon even though I charge it overnight.

    মই ৰাতিৰাতি চাৰ্জ কৰিলেও মোৰ ফোন বেটাৰিৰ মৃত্যু হয়।

How it compares

chrF++ (higher is better) counts how many character sequences and words of a translation match a reference translation. NFC on both sides, our harness, the same sentences for every system. Ours with greedy decoding; indicMT int8 with beam 5, as its card says. Each gap is paired: 1,000 bootstrap resamples of the sentences, 95% interval. Firefox’s offline models greedy, as they ship. Measured 3, 12, 14 and 16 November 2026.

chrF++ for English into each language, higher is better, with the paired 95% intervals of the gaps to indicMT and to Firefox’s offline model
Direction · setPaar (17.3M)indicMT (153.8M)Gap [95% interval]Firefox offlineGap [95% interval]IndicTrans2 dist-200M (211.8M)
English → Hindi · FLORES+55.4956.92−1.43 [−1.89, −0.97]58.13 (17.1 MB)−2.64 [−3.07, −2.16]58.60
English → Hindi · IN22-Conv44.9247.22−2.30 [−2.83, −1.81]46.79 (17.1 MB)−1.87 [−2.36, −1.37]48.79
English → Bengali · FLORES+50.4150.99−0.58 [−1.04, −0.12]51.34 (17.1 MB)−0.93 [−1.41, −0.43]53.11
English → Bengali · IN22-Conv44.8847.21−2.33 [−2.93, −1.76]43.57 (17.1 MB)+1.32 [+0.72, +1.97]48.90
English → Marathi · FLORES+48.0248.68−0.66 [−1.15, −0.19]51.07 (31.6 MB)−3.05 [−3.50, −2.62]50.79
English → Marathi · IN22-Conv44.5546.07−1.52 [−2.04, −0.96]47.70 (31.6 MB)−3.14 [−3.66, −2.60]48.68
English → Telugu · FLORES+55.5056.56−1.06 [−1.56, −0.59]56.62 (17.1 MB)−1.12 [−1.60, −0.61]58.63
English → Telugu · IN22-Conv40.9442.62−1.68 [−2.18, −1.19]40.21 (17.1 MB)+0.73 [+0.26, +1.22]44.69
English → Tamil · FLORES+53.8653.95−0.08 [−0.55, +0.41]56.67 (31.6 MB)−2.81 [−3.24, −2.37]56.67
English → Tamil · IN22-Conv36.1238.14−2.02 [−2.43, −1.62]37.11 (31.6 MB)−0.98 [−1.38, −0.61]38.55
English → Gujarati · FLORES+53.2153.95−0.74 [−1.22, −0.28]52.55 (17.1 MB)+0.66 [+0.13, +1.20]55.62
English → Gujarati · IN22-Conv49.6050.12−0.52 [−1.06, −0.02]47.49 (17.1 MB)+2.11 [+1.59, +2.67]52.94
English → Urdu · FLORES+48.2449.15−0.91 [−1.34, −0.49]50.68 (31.6 MB)−2.44 [−2.81, −2.04]50.91
English → Urdu · IN22-Conv56.0657.91−1.84 [−2.48, −1.18]59.62 (31.6 MB)−3.56 [−4.19, −2.92]60.51
English → Kannada · FLORES+51.9351.65+0.28 [−0.17, +0.71]53.14 (17.1 MB)−1.21 [−1.61, −0.77]54.84
English → Kannada · IN22-Conv30.2931.66−1.37 [−1.68, −1.07]30.64 (17.1 MB)−0.34 [−0.63, −0.01]32.38
English → Malayalam · FLORES+52.6452.22+0.42 [−0.07, +0.94]53.38 (17.1 MB)−0.74 [−1.28, −0.17]55.71
English → Malayalam · IN22-Conv35.5938.22−2.63 [−3.08, −2.22]35.72 (17.1 MB)−0.13 [−0.54, +0.30]38.85
English → Punjabi · FLORES+49.4151.89−2.48 [−2.88, −2.05]none–52.73
English → Punjabi · IN22-Conv51.4754.72−3.25 [−3.89, −2.68]none–55.32
English → Odia · FLORES+46.4050.24−3.85 [−4.35, −3.29]none–47.63
English → Odia · IN22-Conv36.3741.19−4.82 [−5.26, −4.30]none–34.80
English → Assamese · FLORES+39.9041.42−1.51 [−1.94, −1.09]none–42.05
English → Assamese · IN22-Conv42.5744.70−2.13 [−2.72, −1.59]none–42.70

Firefox’s offline models: 9 of the twelve directions, from its model registry (read 14 and 16 November 2026), size as downloaded. Firefox has no Punjabi, Assamese or Odia model.

IndicTrans2 dist-200M is its score as the teacher (beam 4), without a paired test. Parameter counts are ours; Hugging Face lists 274.6M for IndicTrans2 dist-200M, counting embeddings. On Odia IN22-Conv the Odia model scores above IndicTrans2 dist-200M on this one set; we claim nothing from it.

Where it falls short

  • Behind the bigger models: indicMT (153.8M) is ahead on every set we scored except FLORES+ for Kannada, Malayalam and Tamil, where we found no measurable difference, by up to 4.8 chrF++. IndicTrans2 dist-200M, its teacher, is ahead by 0.1 to 4.5 on every set but one. If you can run a 150–200M model, those translate better.
  • Behind Firefox’s offline models in eight languages: Hindi, Marathi, Tamil and Urdu on both test sets, and Bengali, Telugu, Kannada and Malayalam on FLORES+ (Kannada on conversational sentences too), by 0.3 to 3.6 chrF++. Firefox’s Marathi, Tamil and Urdu models are 31.6 MB, 1.6 times ours.
  • Some letters are missing from the vocabulary, which has no byte fallback. They come out as “⁇”, or are dropped by the Transformers decoder. 0.2% to 4.0% of FLORES+ outputs carry one (the letters for each language are in the table below).
  • Names are often mangled, and Latin letters inside names often become “⁇”.
  • Idioms and short phrases are often translated word for word, and small words such as “by” (a deadline) can take the wrong sense.
  • Long formal sentences lose meaning more often than short everyday ones. One sentence at a time: inputs are cut at 127 pieces.
  • Weak spots: conversational Kannada, Malayalam, Tamil and Odia (30.3 to 36.4 on IN22-Conv), and Assamese overall (39.9 to 42.6).
  • English into these twelve only. Translating into English is not part of this release.
  • No native speaker has reviewed the outputs yet.
  • Speed is measured on one Mac mini only (its CPU, and Chrome and Safari), not on phones or other computers.
Letters missing from each model’s vocabulary, an example, and the share of FLORES+ outputs that carry one
LanguageMissingOutputs
Hindiऋ, ॐ, the candra signs ऍ and ॅ, and Devanagari digits (it writes 0–9 instead)2.0%
Bengaliঈ, ঊ, ঋ and ঔ1.7%
Marathiऋ, ऍ, ॐ and the nukta1.6%
Teluguఊ, ఋ, ఔ, ఙ, ఝ and Telugu digits2.4%
Tamilஔ, ௌ, ஶ, ௐ and Tamil digits1.3%
Gujaratiઋ, ઐ, ઔ, ઑ (as in ઑફિસ), ઍ, ઙ, the nukta and ૐFor example, “formal studies” comes out as પચારિક અભ્યાસો, with the first letter of ઔપચારિક missing.2.2%
Urduۂ, ۓ, the hamza mark ٔ and most short-vowel marks (rare in everyday Urdu text)0.2%
Kannadaಙ, ಱ and Kannada digits0.8%
Malayalamഊ, ഋ, ഛ, ഝ, ഢ and Malayalam digits2.4%
PunjabiGurmukhi digits (it writes 0–9 instead), ਙ, ਞ and ੴFor example, a long sentence can end in a run of ⁇ marks when it runs out of words it knows.1.0%
Odiaଈ, ଊ, ଐ and ଔFor example, ଔଷଧ (“medicine”) comes out as ଷଧ.2.9%
Assameseঈ, ঊ, ঋ, ঐ and ঔFor example, ঔষধ (“medicine”) comes out as ষধ.4.0%

Share of FLORES+ devtest outputs with at least one “⁇”, measured 16 November 2026. Check names and numbers in every output.

Run it in Python or JavaScript

Transformers, with the weights as trained:

# pip install transformers sentencepiece torch
from transformers import MarianMTModel, MarianTokenizer

repo = "udgamlabs/Paar-en-gu"
tok = MarianTokenizer.from_pretrained(repo)
model = MarianMTModel.from_pretrained(repo)

batch = tok(["Can you send me the invoice for the last three months?"], return_tensors="pt", padding=True, truncation=True, max_length=128)
out = model.generate(**batch, num_beams=1, max_new_tokens=200)
print(tok.batch_decode(out, skip_special_tokens=True)[0])
# શું તમે મને છેલ્લા ત્રણ મહિના માટે ઇન્વૉઇઝન મોકલી શકો છો?

CTranslate2 int8, the small, fast build in each repo’s ct2-int8 folder:

# pip install ctranslate2 sentencepiece huggingface_hub
import ctranslate2, sentencepiece as spm
from huggingface_hub import snapshot_download

path = snapshot_download("udgamlabs/Paar-en-gu", allow_patterns=["config.json", "ct2-int8/*"])
sp = spm.SentencePieceProcessor(model_file=f"{path}/ct2-int8/source.spm")
tr = ctranslate2.Translator(f"{path}/ct2-int8", device="cpu", compute_type="int8")

def translate(sentences):
    pieces = [p[:127] + ["</s>"] for p in sp.encode(sentences, out_type=str)]
    out = tr.translate_batch(pieces, beam_size=1, max_decoding_length=200)
    return [sp.decode([t for t in r.hypotheses[0] if t != "</s>"]) for r in out]

print(translate(["Can you send me the invoice for the last three months?"])[0])

transformers.js, in a web page or Node, with the int8 ONNX build in each repo’s onnx folder (28.7 MB a language, downloaded once):

// npm install @huggingface/transformers
import { pipeline } from "@huggingface/transformers";

const translate = await pipeline("translation", "udgamlabs/Paar-en-gu", { dtype: "q8" });
const [out] = await translate("Can you send me the invoice for the last three months?", { max_new_tokens: 200 });
console.log(out.translation_text);
// શું તમે મને છેલ્લા ત્રણ મહિના માટે ઇન્વૉઇઝન મોકલી શકો છો?
  • Change -en-gu to the code of another language for the other eleven: -en-hi (Hindi), -en-bn (Bengali), -en-mr (Marathi), -en-te (Telugu), -en-ta (Tamil), -en-ur (Urdu), -en-kn (Kannada), -en-ml (Malayalam), -en-pa (Punjabi), -en-or (Odia), -en-as (Assamese).
  • One sentence per input. Split paragraphs into sentences first.
  • Greedy decoding (num_beams=1) is what we scored. Beam search was not tested.
  • The CTranslate2 int8 build scores within 0.07 chrF++ of the fp32 weights, with 69% to 80% of translations word for word the same; the ONNX int8 build is within 0.16. transformers.js shows a missing letter as <unk> where the Python builds show “⁇”.

To train one further on your own sentence pairs, the GitHub repository has a fine-tuning recipe with Transformers, which also converts the result to int8. If you publish a fine-tuned version, set base_model in its card to the repo you started from.

How it was trained

  1. 1

    English sentences

    From IndicCorp v2 (AI4Bharat, CC0), cleaned and deduplicated; about a third rewritten in a spoken style. Every sentence overlapping FLORES+, NTREX or IN22 was removed before training.

  2. 2

    Translations from IndicTrans2

    IndicTrans2 dist-200M (AI4Bharat, MIT) translated them with beam 4. Those translations are the training targets: the models are distilled from IndicTrans2.

  3. 3

    Training

    A Marian encoder-decoder (6 encoder and 2 decoder layers, width 256, one shared 32,000-piece SentencePiece vocabulary), trained from random weights, one model per direction.

  4. 4

    Scoring

    Scored once on the test sets after training: chrF++, NFC on both sides, greedy decoding, our harness, the same sentences for every system.

Good fits

  • Translating short English text where a large model is too heavy or the text should stay on the device.
  • Apps, offline tools and browser extensions.
  • A first draft for a human translator.

Not for legal, medical or official text without a human check.

Licence

Apache-2.0, for the weights. Copyright 2026 Udgam Labs LLP.

IndicTrans2 dist-200M by AI4Bharat (MIT) wrote the training targets, and the English sentences come from IndicCorp v2 by AI4Bharat (CC0). No third-party weights are included. Details in each model’s NOTICE file. Thank you to AI4Bharat for both.

How to cite

@misc{udgamlabs2026paar,
  title  = {Paar: 17 MB translators from English into 12 languages},
  author = {Prasad, Jitesh K},
  year   = {2026},
  howpublished = {Udgam Labs},
  url    = {https://udgamlabs.com/models/paar}
}

Please also cite IndicTrans2 (Gala et al., 2023, TMLR; arXiv:2305.16307), whose translations it learned from.