At the India AI Impact Summit in February 2026, three AI models made headlines. Turns out, they are just the tip of the iceberg.
When Sarvam, Gnani.ai, and BharatGen unveiled their sovereign models earlier this year at the India AI Impact Summit, it felt like the country’s AI (artificial intelligence) moment had finally arrived. Except, here’s the thing: these three models aren’t the whole story, and they never were. There’s a larger and much wider group of Indian-built AI models that are up and running today.
While some are being worked out of research labs, enterprise war rooms, and other institutes, some of them are already running at population scale.
Think a unicorn-status LLM (large language model) trained from scratch on Indian data, a translation engine covering scripts most of the world has never heard of, an AI model built to understand a farmer’s voice, and more. So, who else is quietly building India’s AI stack? Here’s a roundup of India’s indigenous AI models.
The General-Purpose Contenders
First up, we have to talk about India’s very first homegrown LLM Krutrim, designed and built by Ola’s founder Bhavish Aggarwal. Not only was it the first Indian AI startup to hit unicorn status, all within a few months of launch, but also, it was trained from scratch on Indian data, instead of being a fine-tuned version of an existing foreign model. It understands roughly 22 Indian languages and can generate text in 10 of them.
Today, it’s transformed into a full Krutrim Cloud stack, offering GPU compute (graphics processing unit) paired with consumer assistant Kruti, that can do everything from ordering your dinner to booking your cab.
Then, there’s BharatGPT, which is a collaboration among CoRover, Reliance, and a nine-IIT consortium, which has taken the opposite approach: infrastructure first, headlines later. It might sound unglamorous, but the collective’s Hanooman family of AI models powers IRCTC’s AI assistant. Trained on 22-35 Indian languages, depending on the version, it handles millions of queries a day in Hindi as well as other regional languages.
On the health and enterprise side, it’s led by SML (Seetha Mahalaxmi Healthcare), which is aimed at banking and government services, which in itself earns the idea bragging rights.
The Backbone Nobody’s Talking About
According to the 2011 census, India is a country of 121 languages, 1,300 mother tongues, and a whopping 19,500 localised dialects and sub-dialects. However, none of these flashy sovereign LLMs would function without the all-important language layer underneath them. That would be the efforts of AI4Bharat, a research initiative out of IIT Madras, which has been quietly building the backbone of indigenous AI. Its third-generation IndicTrans series translates across all 22 scheduled Indian languages and spans 13 scripts, a scale that no other open-source can currently match.
Furthermore, AI4Bharat also designed the instruction-tuned Hindi model Airavata by fine-tuning Sarvam’s earlier OpenHathi base, and has released open datasets like IndicCorp and Sangraha that other Indian labs now train on.
All this research feeds straight into the Indian government’s public translation and speech API Bhashini, enabling any app to bolt on multilingual support without having to build a language stack from scratch. And when you pair that with the IndiaAI Mission’s dataset repository AIKosh, what you have is the infrastructure that most of India’s sovereign headline-making LLMs are standing on.
IndiaAI Mission – The Others
Besides Sarvam, Gnani.ai, and BharatGen, 12 other consortia and organisations have been shortlisted to build indigenous AI foundation models. This list includes Soket AI Labs, Gan AI, Avataar AI, an IIT Bombay-led consortium, GenLoop, Zenteq, Intellihealth, Shodh AI, Fractal Analytics, and Tech Mahindra’s Makers Lab, among others.
Of these, it’s worth watching Soket AI closely, as it’s already shipped two models, Pragna and Dhrith. While the former is a compact open multilingual model, the latter is more intriguing: it’s a speech-recognition model designed specifically for Indian accents, solving issues such as emotional tone and mid-sentence language switching, which most global speech models simply aren’t designed to solve.
While Tech Mahindra’s former Project Indus work on Hindi and its dialects now feeds into its broader Makers Lab efforts, Shodh AI and Intellihealth are all about healthcare-specific apps. That’s the exact kind of AI model specific to the Indian context that will really matter.
The Sector Specialists
Besides general-purpose models, there are also sector specialists that are making their mark. Kissan AI’s Dhenu model, for instance, was designed for farmers, working across more than 20 languages. It answers agricultural questions and doubts in places where internet access and formal literacy are both limited.
Zoho’s Zia works differently, integrating AI directly into its existing enterprise software suite. If that wasn’t enough, Krutrim’s chip ambitions and Reliance Jio’s JioBrain show that there’s a parallel race taking place: we aren’t just trying to build the LLMs, we’re also trying to design and build the cloud and chips capacity to train and run them domestically, cutting reliance on global supply chains.
The Voice-First Push
Voice is where most of India lives online, rather than text. Besides Soket AI’s Dhrith and Gnani.ai’s TTS (text-to-speech) model Vachana, there are also a bunch of smaller voice-agent startups building on Bhashini’s speech APIs to power IVR (interactive voice response) systems and customer support in regional languages.
In a country where English fluency and formal literacy exist only in a small percentage, text-only models, by design, lock out a sizeable chunk of the population. Agricultural advisory lines, banking IVR, and government helplines are the first in line to possibly benefit from this, as they already run over phone calls rather than apps.
What Lies Ahead
Two things are obvious here: firstly, most sovereign AI models are optimising for narrower, genuinely harder problems such as language, lower digital literacy, and patchy connectivity, and that too at government scale. Secondly, they’re all talking to each other; BharatGPT is using Bhashini’s API and Bhashini itself is being fed by AI4Bharat’s open research.
With IndiaAI Mission’s funding now available to a dozen more teams whose outputs aren’t even public yet, the hope is that these models do what foreign LLMs have consistently failed to do in India: get to the people who need them most.
In case you missed:
- All About India’s Indigenous AI LLM Models – Can They Help Tackle Bias?
- All About AI Sovereignty: Why It Matters To India
- All About AI Sandboxing
- Are You Using AI To Generate Passwords? It’s A Terrible Idea
- AI Infrastructure: Building Scalable Architecture For The Future
- AgentOps: The Dawn Of The Internet Of Agents
- The Rise and Evolution Of Honeypots In Cybersecurity
- Agentic AI Is Now A Fund Manager, Believe It Or Not
- All About Data Poisoning Cyberattacks
- AI Runtime Security: A Comprehensive Look
