India seems to be figuring out the development of sovereign LLMs—with the emphasis firmly on “sovereign”. With homegrown initiatives like Indus, Sarvam, Krutrim, and regional LLMs undergoing training, such as Soket AI Labs, suggest that the country now understands the pipelines, at least for the next six to 10 months.
The IndiaAI Mission has laid out support and foundation for startups to build the next LLM. However, some argue that much of the government-backed AI development has focused on proving that India can build LLMs, rather than precisely defining what they are being built for.
“We know the art of building LLMs. Now it’s time to move on,” Nikhil Malhotra, chief innovation officer at Tech Mahindra and the person behind Project Indus, told AIM. According to him, the next frontier isn’t in model architecture or token counts but in building for India’s fragmented, multilingual and real-world problems.
Though the four selected startups, Sarvam AI, Soket AI, Gnani.ai, and Gan.AI, have each chosen their areas of focus, there remains a broader need to identify more use cases and, possibly, build models that solve those issues.
For instance, Indian companies are still turning to OpenAI APIs or Meta’s Llama for a simple reason: Indian LLMs often aren’t ready for production. One such example is Zepto’s latest blog post about using Llama 3 for multilingual capabilities on their search bar.
Latency issues and a lack of fine-tuning of Indian models make them hard to deploy. However, this should not be seen as a shortcoming of the models themselves, as Malhotra pointed out. “They were built as base models. What we’re missing is distillation and sector-level fine-tuning.”
When Sarvam released Sarvam-M, its first LLM with IndiaAI Mission, there was massive support and also severe backlash from the community as it was built on top of Mistral. Deedy Das from Menlo Ventures called the number of downloads “embarrassing” and said that there is no need to build Indic LLMs without identifying use cases.
Read: Sarvam AI’s Backlash Exposes the Sad State of Indian AI
What the IndiaAI Mission Should Also Fund
Echoing Das’ perspective, Malhotra said, “There’s no need to reinvent the wheel. The government should back efforts solving what no one else will—dialects, local problems, fringe cases that are mainstream in India.”
Instead of building yet another monolithic LLM, India’s focus should shift to domain-specific, distilled models tailored for applications: chatbots in banks, dialect-aware agricultural assistants, or customer support systems across Indic languages.
Malhotra believes the IndiaAI Mission’s funding should be split 70% between refining and deploying existing models into public systems, and 30% towards truly new research. That research could focus on multilingual multimodal AI, AI-native operating systems like Bharat Operating System Solutions, or education and agriculture-specific platforms.
Along similar lines, Vivekanand Pani, co-founder and CTO of Reverie Language Technologies, said that he believes India’s race to build LLMs is beginning to look more like a panicked response than a strategic mission.
The country’s ₹10,000 crore AI initiative is centred on building foundational models within ten months. Pani argues that this “reactive” mindset might be solving the wrong problem altogether. India, he said, is trying to build use cases by first building models, when it should be the other way around.
Stop Building the Bomb. Start Asking Who You’re Trying to Protect.
Use cases naturally emerge when people actively engage with the internet in their own language for everyday tasks. “Imagine the US builds the first nuclear bomb,” Pani told AIM, while sharing a powerful analogy. “No one understands its true power.”
“The moment it is used, the world wakes up and starts reacting. However, no one gets the blueprint. No one shares the uranium. Countries have to figure that out themselves.” In the AI world, he compared OpenAI’s release of ChatGPT to this moment. Since then, India has been scrambling to catch up, without asking what it really wants to do with the technology.
“Are we building another nuclear bomb just because the US did? Or do we need a reactor to generate energy? We haven’t chosen our target. We haven’t even confirmed if we have uranium,” Pani said. In this case, by uranium, he meant the Indic data needed to build effective models for Indian users. Notably, it’s in short supply.
“Use cases aren’t imagined. They emerge from lived engagement. If we help people use the internet in Indian languages, the use cases will show up naturally,” he said.
Read: Indic AI is Not Inspiring Enough for Indian Developers
According to Pani, the government should be catalysing engagement, not throwing money at building LLMs without a clear sense of how they will be used. “What we needed was not an AI model, but a national mission for digital language engagement—in schools, in governance, in business, in agriculture,” he said.
A Case of ‘Reactionary LLMs’
Pani also pointed to the release of Sarvam-M as an example of reactionary development. “Why build on Mistral if we’re talking about sovereignty? Sarvam, too, started with the idea that ‘we need a nuclear bomb’ because everyone else has one. But we haven’t asked ‘for what purpose?’”
To his credit, Pani acknowledges that Sarvam’s effort may still lead to algorithmic innovation. “They started work on tokenisation efficiency, which will matter if they eventually build a foundational model. But without engagement, none of this turns into usable solutions,” he said.
The real issue, he argues, is that we’re trying to solve the problem of LLM parity with the West, when we haven’t even solved the problem of language access for our own people. “I don’t care who builds the tech. I just want it to enable engagement in schools, villages and homes. That’s what we missed. That’s the foundation,” he said.
Rather than attempting to create a universal Indian LLM, Malhotra also envisions a decentralised approach. Multiple LLMs, each fine-tuned for language and region, could be orchestrated through a common platform.
“Think of it as an LLM switchboard: if a user speaks Telugu, route the query to the Telugu model. If it’s in Odia, switch to the Oriya LLM,” he said. “We need orchestration, not replication. Use the diversity that already exists. The job is to reduce latency, distil the models, and make them work in use cases startups can plug into immediately.”
He emphasised starting with freemium deployments—as OpenAI did with ChatGPT—to drive adoption. “No one will pay until they see it working.”
Startups, in this framing, should shift from building their own LLMs to innovating on top of open models. That includes distilling large models for mobile deployment, domain-specific fine-tuning, and building sectoral applications.
One of the biggest issues that Indian AI startups are facing is the lack of enterprise support. Companies frequently demand POCs but hesitate to pay. Malhotra suggests that the government can step in—not just with funding but with legal and operational backing that gives AI startups room to persist.
“The government should back Indian AI not just with money, but with belief. And that belief should extend to giving wings to those solving problems others won’t touch,” he added.
While global models like Llama and Gemini may now support more Indian languages, sovereignty is about more than language support. It’s about control over biases, ethics, and data residency.
“India’s fringe is the West’s norm. You can’t solve Indian problems with Western models alone,” Malhotra said. Startups must understand they won’t win on model size or compute scale. Their edge lies in vertical focus, creative deployment, and solving India-specific edge cases.
The post Sovereign LLMs Won’t Alone Fix the Broken Indian AI Ecosystem appeared first on Analytics India Magazine.
