I’ve been saying it for months: India isn’t just a consumer of AI anymore—we’re officially building the engine. At the India AI Impact Summit in New Delhi, the government and three home-grown startups just dropped a trio of “sovereign” AI models that are designed to do something Big Tech often struggles with—actually understanding the nuances of India.
If you’ve felt like ChatGPT or Gemini don’t quite “get” the local context, here is why this week’s news changes everything.
The Big Picture: India’s ₹10,000 Crore Bet
To be honest, we’ve been waiting for the IndiaAI Mission to show its teeth since it was cleared back in 2024. Now, with over ₹100 crore already pushed into GPU subsidies and compute support, we are seeing the first real results.
-
The Goal: Build a domestic “foundational model” so we aren’t dependent on foreign tech.
-
The Scale: Over 38,000 GPUs are now being deployed to power these indigenous systems.
Why it matters: Sovereignty isn’t just a buzzword. It means our data stays here, our laws govern the AI, and the costs are optimized for our economy, not Silicon Valley’s.
Sarvam AI: The Heavy Hitters (30B & 105B Models)
Bengaluru-based Sarvam AI didn’t just show up; they brought the heat. They unveiled two Large Language Models (LLMs) that are built from scratch on Indian data.
-
Sarvam-30B: Small, fast, and incredibly efficient.
-
Sarvam-105B: The flagship. It supports a massive 128,000-token context window.
-
The Performance: Sarvam claims their larger model is outperforming Google’s Gemini Flash and DeepSeek R1 on several benchmarks while being cheaper to run.
Why it matters: Most AI models “hallucinate” or get confused when you mix Hindi and English (Hinglish). Sarvam was trained on 16 trillion tokens of Indian data, meaning it actually understands the way we talk.
Gnani.ai: The Voice of 12 Indian Languages
Here’s the thing: most of India interacts with technology through voice, not text. Gnani.ai’s new Vachana TTS is a game-changer for accessibility.
-
Feature: Can clone a human voice using less than 10 seconds of audio.
-
Language Support: 12 Indian languages, including Hindi, Tamil, and Kannada.
-
Tech: Built for low-bandwidth areas (think rural India with spotty 4G).
Why it matters: This isn’t just for fancy apps. Imagine a government service or an emergency alert system that speaks to you in a voice that sounds familiar and natural, even if you’re in a remote village with a basic feature phone.
BharatGen: AI for the Common Good
Led by IIT Bombay, BharatGen is focusing on the “Sovereign” part of the mission. Their Param2 17B model is the largest beneficiary of the IndiaAI Mission so far.
-
Model: 17-billion-parameter Mixture-of-Experts (MoE) system.
-
Accessibility: It’s Open Source. You can literally find it on Hugging Face right now.
-
Focus: Governance, healthcare, and education.
Why it matters: By making this open-source, BharatGen is letting every Indian startup and researcher build on top of their foundation for free. It’s the “Android” approach to India’s AI ecosystem.
My Take: Sundar Pichai said it best at the summit: the “developer energy” in India is second to none. We are finally moving from being a “service economy” to a “product economy.” These models are the first step tow
