Indic LLM Platform

In plain words

IndicTrans2 is a machine translation model that converts text between English and all 22 official Indian languages, and between Indian languages directly. It is built to give high-quality translations even for languages with little online data. It is open source and free to run on your own computer or servers.

How to use it

🚧

This model is coming soon!

A step-by-step guide for IndicTrans2 will appear here once it goes live.

Languages & scripts supported

Works in these languages, in both native script and Roman typing.

English Englishঅসমীয়া Assameseবাংলা Bengaliबड़ो Bodoडोगरी Dogriગુજરાતી Gujaratiहिन्दी Hindiಕನ್ನಡ Kannadaکٲشُر Kashmiriकोंकणी Konkaniमैथिली Maithiliമലയാളം Malayalamমৈতৈলোন্ Manipuri (Meitei)मराठी Marathiनेपाली Nepaliଓଡ଼ିଆ Odiaਪੰਜਾਬੀ Punjabiसंस्कृतम् Sanskritᱥᱟᱱᱛᱟᱲᱤ Santaliسنڌي Sindhiதமிழ் Tamilతెలుగు Teluguاردو Urdu+ Roman & code-mixed

Strengths & limits

An honest look at what it does well and where it struggles.

Good at

✓High-quality Indian language translation

✓English to and from Indian languages

✓Direct Indian-to-Indian language translation

✓Low-resource and regional languages

Where it struggles

!Translation only, not a chatbot

!Cannot answer questions or chat

!Quality varies by language pair

!Needs setup to self-host

Technical details & licence

Training dataTrained on the Bharat Parallel Corpus Collection (BPCC), a large dataset of around 230 million sentence pairs. It includes mined web text plus human-translated and quality-checked sentences across all 22 languages.

SizeFull models are about 1 billion parameters (En-Indic, Indic-En, Indic-Indic), with smaller distilled versions around 200M for faster, lighter use. Newer long-context variants handle sequences up to 2048 tokens.

LicenceModel checkpoints are released under the MIT licence, allowing free commercial and research use. Training data uses open licences (CC0 and CC-BY-4.0).

Commercial costFree to self-host

Version historyIndicTrans2 · May 2023 — first model to cover all 22 scheduled Indian languagesDistilled · 2023 — smaller ~200M variants for faster, cheaper useLong-context · Jan 2025 — RoPE-based variants for longer text up to 2048 tokens

About the maker

Who builds and maintains IndicTrans2.

AI4Bharat

4 models

AI4Bharat is an open-source research lab at IIT Madras dedicated to advancing AI for Indian languages. Its freely released datasets, benchmarks, and models — like Airavata and the IndicTrans translation series — are widely used across research and industry.

View their models →See all organisations →