News
ReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior ReleaseReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior Release
EveryLLM
IndicTrans2

IndicTrans2

Machine translation for 22 scheduled Indian languages

IndicTrans2 is a machine translation model supporting all 22 scheduled Indian languages. Released in May 2023 by researchers including Mitesh M. Khapra and Anoop Kunchukuttan, it addresses the lack of high-quality translation tools for India's linguistic diversity. The project includes the Bharat Parallel Corpus Collection and open-access models.

Share on WhatsApp

  • Supports all 22 scheduled Indian languages
  • Includes 230 million bitext pairs in BPCC
  • Released with permissive open licenses
  • First n-way parallel benchmark for Indic languages
  • Surpasses existing models on new benchmarks

What it is

IndicTrans2 is a machine translation system designed to handle all 22 scheduled languages of India. It was developed to fill gaps in existing technology, as no prior models supported this full set of languages with high quality. The project focuses on accessibility and performance, providing tools for a population speaking languages from four major families. It serves as a resource for developers and researchers working with Indian linguistic content.

How it came about

The work was submitted to arXiv in May 2023 and revised in December 2023. Authors include Jay Gala, Pranjal A. Chitale, Raghavan AK, and Mitesh M. Khapra, among others. The team identified three main missing pieces in the field: parallel training data for all 22 languages, robust benchmarks with Indian content, and models supporting these languages. They addressed these by curating new datasets and training multilingual models.

Inside the model

The system relies on the Bharat Parallel Corpus Collection (BPCC), which contains 230 million bitext pairs. Of these, 126 million were newly added, including 644,000 manually translated sentence pairs. The model is multilingual, trained to translate between any of the 22 scheduled languages. The architecture supports n-way parallel translation, allowing diverse domain coverage and Indian-origin content in its benchmarks.

What it can do

IndicTrans2 surpasses existing models on multiple benchmarks, including those created specifically for this work. It supports translation across all 22 scheduled Indian languages, a capability not found in prior systems. The associated benchmark features diverse domains and source-original test sets, ensuring relevance to local contexts. Performance improvements are noted in both existing and new evaluation metrics.

How to use it

The models and associated data are released with permissive licenses to promote collaboration. This open access approach aims to enable wide and easy use of high-quality translation systems. Researchers and developers can download the weights and use the Bharat Parallel Corpus Collection for further training or evaluation. The goal is to make these resources accessible to the broader community.

Why it matters

This work provides the first n-way parallel benchmark covering all 22 Indian languages. It establishes a baseline for future research in Indic language processing. By releasing the largest publicly available parallel corpora for these languages, it facilitates the development of more accurate and inclusive translation tools. The project highlights the importance of localized data in building effective AI systems for diverse linguistic regions.

Questions people ask about IndicTrans2

Which languages does IndicTrans2 support?

It supports all 22 scheduled languages listed in the Constitution of India. These languages belong to four major language families and are spoken by over a billion people. The model is designed to handle translation between any pair of these languages.

What data was used to train the model?

The training used the Bharat Parallel Corpus Collection (BPCC). This dataset contains 230 million bitext pairs. The contributors added 126 million new pairs, including 644,000 that were manually translated to ensure high quality and relevance.

Is the model available for public use?

Yes, the models and data are released with permissive licenses. This open access strategy aims to promote collaboration and make high-quality machine translation accessible to developers and researchers working with Indian languages.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on IndicTrans2 yet, so the facts on this page come from the sources listed above: arxiv.org/abs/2305.16307. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.

    It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

    The weekly digest

    Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.