BharatGen

Text, speech and vision AI models for India

Four families, each pretrained by our own team on Indian data, with open weights on Hugging Face and AIKosh.

Our model families

Text · LLM

Param

Large language models for 22 Indian languages and English. Param2, a 17 billion parameter mixture of experts, reasons, codes and uses tools; domain models cover Ayurveda, agriculture, finance and law.

Param2 17BParam1AyurParamAgriParamFinanceParamLegalParam
Text models → Vision · VLM

Patram

Vision language models built for India. Patram reads Indian documents and images, answers questions about them, extracts key fields and reasons over tables and charts.

Patram 1.2Patram 1.1Patram 1
Vision models → Speech · Speech-to-text

Shrutam

Speech-to-text for 12 Indian languages, built on a language model, so it keeps up with code-switching in real time.

Shrutam 2Shrutam 1 Hindi
Speech models → Speech · TTS

Sooktam

Text-to-speech in 11 Indian languages and Indian English, with zero-shot voice cloning and cross-lingual voices.

Sooktam 2
Speech models →

Every release so far

From our first models in June 2025 to our second generation in February 2026.

  1. Patram 1, Shrutam 1, Sooktam 1Vision and speech
  2. Param1Text
  3. Patram 1.1Vision
  4. AyurParam, AgriParam, FinanceParam, LegalParamText, by domain
  5. Param1 5B, Param1 7BText
  6. Param2, Shrutam 2, Sooktam 2, Patram 1.2All four families

How to use our AI models

Everything you need is on the model cards.

Step 1

Pick a model

Each model page lists what it does, its languages and its size. Start small on a laptop.

Step 2

Download the weights

From Hugging Face or AIKosh. Some ask you to accept their terms first.

Step 3

Run and adapt

Use the code on the model card, then fine-tune on your own data.