BharatGen logo

BharatGen models

Government-funded AI initiative anchored at IIT Bombay, run since November 2025 by the not-for-profit BharatGen Technology Foundation, that pretrains open-weight text, speech and document-vision models for Indian languages.

Mumbai, IndiaFounded 2024bharatgen.com

Releases
13
Open weights
13
Latest
Feb 17, 2026

February 2026

3 releases
  1. ReasoningOpen weights

    Param2

    Param2 is BharatGen's second-generation language model, a mixture-of-experts model with 17 billion parameters pretrained from scratch on Indian-language and English data. The released checkpoint is a reasoning model that writes out its thinking before it answers.

    Modalities
    text → text
    Context
    4K context
    Parameters
    MoE: 17B total, 2.4B active
    License
    BharatGen Non-Commercial License
    At launch
    Open weights
    • Covers English and all 22 scheduled Indian languages with its own 128,000-entry tokenizer vocabulary
    • Routes each token to 6 of 64 experts plus 2 shared experts, so about 2.4B parameters run per token
    • Launch card listed a 4,096-token context and about 22 trillion training tokens; a card update on August 28, 2026 raised the supported context to 32,768 tokens
    • Launched at the India AI Impact Summit as an early post-training checkpoint, with weights on Hugging Face and AIKosh
    Source: Param2: what we built and how | BharatGen
  2. SpeechOpen weights

    Shrutam2

    Shrutam2 is BharatGen's second-generation speech recognition model for 12 Indian languages. It connects a Conformer speech encoder to a pretrained language model decoder through a mixture-of-experts projection layer.

    Modalities
    audio, text → text
    License
    BharatGen Research License
    At launch
    Open weights
    • Covers Hindi, Marathi, Tamil, Telugu, Malayalam, Kannada, Odia, Bengali, Urdu, Assamese, Gujarati and Punjabi
    • A text prompt sets the target language, and BharatGen says it handles speakers who switch languages mid-sentence
    • Launched at the India AI Impact Summit on February 17, 2026
    • Weights followed on Hugging Face on February 23, 2026, under the BharatGen Research License for non-commercial research and academic use
    Source: Speech models | BharatGen
  3. SpeechOpen weights

    Sooktam2

    Sooktam2 is BharatGen's second-generation text-to-speech model, covering 11 Indian languages plus Indian English. Given a short reference clip, it can speak in that voice, including in other languages.

    Modalities
    text, audio → audio
    License
    MIT
    At launch
    Open weights
    • Languages are Hindi, Marathi, Gujarati, Tamil, Telugu, Kannada, Bengali, Malayalam, Odia, Urdu, Punjabi and Indian English
    • Tagged on its model card as an F5-TTS based model, and described by BharatGen as fast enough for live apps and phone calls
    • Launched at the India AI Impact Summit on February 17, 2026
    • Weights followed on Hugging Face on February 24, 2026; the card first listed MIT and switched to the BharatGen non-commercial license on March 9, 2026
    Source: Speech models | BharatGen

December 2025

2 releases
  1. LLMOpen weights

    Param1-5B

    Param1-5B is a dense Hindi-English base model with 5 billion parameters in BharatGen's Param1 family, pretrained from scratch with more math and code data than Param1-2.9B.

    Modalities
    text → text
    Parameters
    5B
    License
    BharatGen Non-Commercial License
    At launch
    Open weights
    • Bilingual in English and Hindi, pretrained from scratch
    • Uses a revised data mix aimed at stronger math reasoning and code generation than Param1-2.9B
    • Released as a pretrained base model for downstream fine-tuning
    • Weights published on Hugging Face and AIKosh under the BharatGen non-commercial license
    Source: bharatgenai/Param-1-5B · Hugging Face
  2. LLMOpen weights

    Param1-7B MoE

    Param1-7B is a mixture-of-experts base model in BharatGen's Param1 family, pretrained from scratch for English, Hindi and 14 other Indian languages and intended as a starting point for fine-tuning.

    Modalities
    text → text
    Context
    4K context
    Parameters
    7B MoE (64 experts)
    License
    BharatGen Non-Commercial License
    At launch
    Open weights
    • Trained on 4 trillion tokens, with 64 experts selected dynamically for each token
    • Adds Assamese, Bengali, Gujarati, Kannada, Maithili, Malayalam, Marathi, Nepali, Odia, Punjabi, Sanskrit, Sindhi, Tamil and Telugu to English and Hindi
    • Released as a pretrained base checkpoint rather than an instruction-tuned model
    • Published on the same day as Param1-5B, under the BharatGen non-commercial license
    Source: bharatgenai/Param-1-7B · Hugging Face

October 2025

2 releases
  1. Small modelOpen weights

    FinanceParam

    FinanceParam is an English-Hindi model with 2.9 billion parameters from BharatGen, fine-tuned from Param1-2.9B-Instruct for questions on Indian tax, banking, investments and financial policy.

    Modalities
    text → text
    Context
    2K context
    Parameters
    2.9B
    License
    BharatGen Non-Commercial License
    At launch
    Open weights
    • Trained on 24 million samples built from more than 25,000 Indian finance passages, including question-answer pairs and multi-turn dialogues in English and Hindi
    • Sources included official material from the Income Tax Department, RBI, SEBI and IRDAI
    • Evaluated on BharatGen's BhashaBench-Finance benchmark
    • Its Hugging Face card moved from the BharatGen non-commercial license to Apache 2.0 on February 24, 2026
    Source: FinanceParam: an AI model for Indian finance | BharatGen
  2. Small modelOpen weights

    LegalParam

    LegalParam is an English-Hindi model with 2.9 billion parameters from BharatGen, fine-tuned from Param1-2.9B-Instruct on Indian law to answer legal questions and summarize legal and policy documents.

    Modalities
    text → text
    Context
    2K context
    Parameters
    2.9B
    License
    BharatGen Non-Commercial License
    At launch
    Open weights
    • Trained on 5 million question-answer pairs: about 2 million curated from open data on Indian Acts and policies plus 3 million synthetic examples
    • Adds special tokens for structured prompts and step-by-step reasoning
    • Evaluated on BharatGen's BhashaBench-Legal benchmark
    • Its Hugging Face card moved from the BharatGen non-commercial license to Apache 2.0 on February 24, 2026
    Source: LegalParam: an AI model for Indian law | BharatGen

September 2025

1 release
  1. Small modelOpen weights

    AyurParam

    AyurParam is an English-Hindi model with 2.9 billion parameters from BharatGen, fine-tuned from Param1-2.9B-Instruct on classical and modern Ayurveda texts.

    Modalities
    text → text
    Context
    2K context
    Parameters
    2.9B
    License
    CC-BY-4.0
    At launch
    Open weights
    • Training data drew on about 1,000 books, including 600 classical manuscripts, covering about 150,000 pages
    • Handles Sanskrit terms and transliteration from texts such as the Charak Samhita and Sushrut Samhita
    • BharatGen reported that it beat similar-sized models on its BhashaBench-Ayur benchmark
    • Its launch card carried a CC BY 4.0 badge; the card listed Apache 2.0 from February 24, 2026
    Source: AyurParam: an AI model for Ayurvedic knowledge | BharatGen

August 2025

1 release
  1. Small modelOpen weights

    AgriParam

    AgriParam is an English-Hindi model with 2.9 billion parameters from BharatGen, fine-tuned from Param1-2.9B-Instruct for farmer advisories, crop and scheme information and rural extension work.

    Modalities
    text → text
    Context
    2K context
    Parameters
    2.9B
    License
    BharatGen Non-Commercial License
    At launch
    Open weights
    • Fine-tuned on about 12 million agriculture samples, including question-answer pairs and simulated advisory conversations in English and Hindi
    • Data follows a taxonomy of crops, weather, soil, pests, fertilizers, markets and government schemes, with personas from smallholder farmers to policymakers
    • Evaluated on BharatGen's BhashaBench-Krishi benchmark
    • Its Hugging Face card moved from the BharatGen non-commercial license to Apache 2.0 on February 24, 2026
    Source: AgriParam: an AI model for Indian farming | BharatGen

July 2025

1 release
  1. Small modelOpen weights

    Param1

    Param1 is BharatGen's first foundation language model, a Hindi-English model with 2.9 billion parameters trained from scratch on 7.5 trillion tokens. BharatGen released an instruction-tuned version on Hugging Face and the pretrained checkpoint on AIKosh.

    Modalities
    text → text
    Context
    2K context
    Parameters
    2.9B
    License
    BharatGen Non-Commercial License
    At launch
    Open weights
    • Hindi made up 2.77 trillion of the 7.5 trillion pretraining tokens, across three phases ending with long-document training
    • BharatGen reported that it led Sarvam-1 on four of the five India-centric benchmarks it tested
    • Released alongside BharatGen-128K, a tokenizer for Indian scripts, though Param1 itself used NVIDIA's Nemotron tokenizer
    • The instruct model's card moved from the BharatGen non-commercial license to Apache 2.0 on February 24, 2026
    Source: Param1: a bilingual AI model built from scratch | BharatGen

June 2025

3 releases
  1. MultimodalOpen weights

    Patram1

    Patram is BharatGen's first vision-language model, a 7-billion-parameter model that reads document images such as forms, bills and certificates and answers questions about them without a separate OCR step.

    Modalities
    image, text → text
    Parameters
    7B
    At launch
    Open weights
    • Trained for Indian document formats including Aadhaar layouts, IFSC codes, cheques, stamps, seals and government forms
    • A government press release described it as trained on 2.5 billion tokens
    • BharatGen describes the Patram releases as English-first, with some ability on Indian scripts
    • Later revisions in the same repository added multi-turn chat and key-value extraction (Patram1.1, August 2025) and reasoning, OCR and chart reading (Patram1.2, February 2026)
    Source: Vision models | BharatGen
  2. SpeechOpen weights

    Shrutam1

    Shrutam1 is BharatGen's first speech recognition model, a compact Hindi speech-to-text system released at the BharatGen Summit in New Delhi.

    Modalities
    audio → text
    Parameters
    30M
    At launch
    Open weights
    • Supports Hindi only
    • A government press release gave its size as 30 million parameters
    • Distributed as an ESPnet model, with inference through a custom ESPnet branch
    • Followed by the 12-language Shrutam2 in February 2026
    Source: Speech models | BharatGen
  3. SpeechOpen weights

    Sooktam1

    Sooktam1 is BharatGen's first text-to-speech model, a 150-million-parameter system trained from scratch on Indian speech with a separate voice model for each language.

    Modalities
    text → audio
    Parameters
    150M
    License
    MIT
    At launch
    Open weights
    • Covers nine Indian languages; the Hindi model was the one published on Hugging Face
    • The Hindi checkpoint was released under the MIT license
    • Released at the BharatGen Summit in New Delhi alongside Patram and Shrutam
    • Followed by Sooktam2 in February 2026
    Source: Speech models | BharatGen

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. "BharatGen models: release history and timeline." The AI Engineering, updated Feb 17, 2026. https://theaiengineering.com/releases/bharatgen

HTML link

<a href="https://theaiengineering.com/releases/bharatgen">BharatGen 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