AI4Bharat logo

AI4Bharat models

Research lab at IIT Madras that releases open datasets and models for Indian languages, spanning translation, transliteration, speech recognition, speech synthesis and language models.

Chennai, Indiaai4bharat.iitm.ac.in

Releases
9
Open weights
9
Latest
Dec 20, 2025

December 2025

1 release
  1. EmbeddingOpen weights

    IndicBERT v3

    IndicBERT v3 is AI4Bharat's third generation of IndicBERT encoders, built by converting Gemma 3 into bidirectional models for language-understanding tasks in Indian languages. It launched in 1B and 4B sizes.

    Modalities
    text → embedding
    Parameters
    1B and 4B
    License
    MIT
    At launch
    Open weights
    • Covers 23 languages: English, 14 major Indian languages and 8 low-resource ones such as Sanskrit and Santali
    • Trained with masked next-token prediction instead of BERT-style masked language modeling, on sequences of up to 4,096 tokens
    • Keeps the standard Gemma 3 vocabulary; the 1B model starts from gemma-3-1b-pt
    • A 270M version followed on Hugging Face on January 6, 2026
    Source: ai4bharat/IndicBERT-v3-1B · Hugging Face

April 2025

1 release
  1. LLMOpen weights

    IndicTrans3-beta

    IndicTrans3-beta is an AI4Bharat translation model built on Gemma 3 and fine-tuned for sentence-level and document-level translation between English and Indian languages.

    Modalities
    text → text
    License
    CC-BY-4.0
    At launch
    Open weights
    • Primary support for 15 Indian languages, including Assamese, Hindi, Maithili, Sanskrit, Tamil and Urdu
    • Preliminary support for seven low-resource languages such as Bodo, Dogri, Santali and Sindhi, with a warning that quality may vary
    • Translates in both directions between English and Indian languages
    • Released under CC BY 4.0 behind a click-through access form on Hugging Face
    Source: ai4bharat/IndicTrans3-beta · Hugging Face

March 2025

2 releases
  1. SpeechOpen weights

    IndicF5

    IndicF5 is AI4Bharat's text-to-speech model for 11 Indian languages. It speaks new text in the voice and style of a short reference recording supplied with its transcript.

    Modalities
    text, audio → audio
    License
    MIT
    At launch
    Open weights
    • Trained on 1,417 hours of speech from Rasa, IndicTTS, LIMMITS and IndicVoices-R
    • Covers Assamese, Bengali, Gujarati, Hindi, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil and Telugu
    • AI4Bharat credits F5-TTS as the inspiration for the work; output is 24 kHz audio
    • MIT-licensed, with terms that limit cloning to voices the user has permission to use
    Source: ai4bharat/IndicF5 · Hugging Face
  2. SpeechOpen weights

    IndicSeamless

    IndicSeamless is AI4Bharat's fine-tune of Meta's SeamlessM4T v2 for speech recognition and speech-to-text translation in Indian languages, trained on the lab's BhasaAnuvaad dataset.

    Modalities
    audio → text
    License
    CC-BY-NC-4.0
    At launch
    Open weights
    • Fine-tuned on BhasaAnuvaad, a speech translation corpus AI4Bharat described in a November 2024 paper
    • Its model card tags list 13 languages
    • Training data was filtered with an alignment score threshold of 0.8 and a mining score threshold of 0.6
    • Released under CC BY-NC 4.0, which bars commercial use
    Source: ai4bharat/indic-seamless · Hugging Face

October 2024

1 release
  1. SpeechOpen weights

    Indic Parler-TTS

    Indic Parler-TTS is a multilingual text-to-speech model from AI4Bharat and the Hugging Face audio team that extends Parler-TTS Mini to Indian languages. It takes the text to speak plus a written description of how the voice should sound.

    Modalities
    text → audio
    License
    Apache-2.0
    At launch
    Open weights
    • Trained on 1,806 hours of Indic and English speech, fine-tuned from the Indic Parler-TTS Pretrained checkpoint
    • Its card said it officially spoke 20 Indian languages plus English, with 69 recommended voices
    • Emotion-specific prompts were officially supported in 10 languages, including Assamese, Bengali, Tamil and Sanskrit
    • Released under Apache 2.0
    Source: ai4bharat/indic-parler-tts · Hugging Face

September 2024

1 release
  1. SpeechOpen weights

    IndicConformer

    IndicConformer is AI4Bharat's suite of open speech recognition models for all 22 scheduled Indian languages, released as one Conformer model per language.

    Modalities
    audio → text
    License
    MIT
    At launch
    Open weights
    • Hybrid CTC and RNN-T Conformer models that run on AI4Bharat's fork of NVIDIA NeMo
    • AI4Bharat called it the country's first open-source ASR system covering so many languages
    • Released under the MIT license
    • A single 600M-parameter multilingual model for all 22 languages followed on Hugging Face on March 15, 2025
    Source: GitHub - AI4Bharat/IndicConformerASR

January 2024

1 release
  1. LLMOpen weights

    Airavata

    Airavata is AI4Bharat's Hindi instruction-tuned language model, made by fine-tuning Sarvam AI's OpenHathi base model on Hindi instruction data.

    Modalities
    text → text
    Parameters
    7B
    License
    Llama 2 Community License
    At launch
    Open weights
    • Fine-tuned with LoRA on the IndicInstruct collection, which AI4Bharat released alongside the model
    • AI4Bharat built the instruction data from human-curated, license-friendly sources and did not use outputs from proprietary models such as GPT-4
    • Released with Hindi evaluation benchmarks and a technical report on arXiv
    • AI4Bharat said it planned to extend support to all 22 scheduled languages
    Source: Airavata: Introducing Hindi Instruction-tuned LLM Model

May 2023

2 releases
  1. LLMOpen weights

    IndicTrans2

    IndicTrans2 is AI4Bharat's open multilingual translation model family covering all 22 scheduled Indian languages, released with its training corpus and benchmarks.

    Modalities
    text → text
    Parameters
    1B base; 200M distilled
    License
    MIT
    At launch
    Open weights
    • Includes 1B base models for English-to-Indic and Indic-to-English translation, plus distilled 200M versions
    • Released with the Bharat Parallel Corpus Collection of about 230 million sentence pairs and the IN22 test set
    • Model checkpoints released under the MIT license
    • Hugging Face-compatible checkpoints followed on September 9, 2023, and Indic-to-Indic models on December 1, 2023
    Source: GitHub - AI4Bharat/IndicTrans2: Translation models for 22 scheduled languages of India
  2. SpeechOpen weights

    IndicWhisper

    IndicWhisper is a set of Whisper models that AI4Bharat fine-tuned for speech recognition in 12 Indian languages, released with the Vistaar benchmark suite.

    Modalities
    audio → text
    License
    MIT
    At launch
    Open weights
    • Fine-tuned on about 10,700 hours of publicly available labeled speech
    • AI4Bharat reported the lowest word error rate on 39 of the 59 Vistaar benchmarks, with an average reduction of 4.1 points
    • Released as separate checkpoints per language, from Bengali to Urdu, under the MIT license
    • The accompanying paper was accepted at Interspeech 2023
    Source: GitHub - AI4Bharat/vistaar: Vistaar: Diverse Benchmarks and Training Sets for Indian Language ASR

Cite and reuse this data

These release records are free to reuse under CC BY 4.0, including commercially. Credit The AI Engineering and link to this page's URL.

Dates are the labs' official announcement dates, and every record links to its source. List fields in the CSV are separated by " | ".

Citation

The AI Engineering Editors. "AI4Bharat models: release history and timeline." The AI Engineering, updated Dec 20, 2025. https://theaiengineering.com/releases/ai4bharat

HTML link

<a href="https://theaiengineering.com/releases/ai4bharat">AI4Bharat models: release history and timeline</a> (The AI Engineering)
Full dataset
1,250 releases from 73 labs
Coverage
to
Updated
License
CC BY 4.0
Publisher
The AI Engineering

Other labs