For all the noise around ‘AI for Bharat’, the developers who are supposed to build it appear to have very little reason to do so. One of the most prominent examples of this was the launch of Sarvam’s AI model, which kick-started the debate around Indic language AI. 

Sarvam-M saw paltry downloads trickle in over a week, prompting many to question the need for building an Indic model in the first place.

Cut to the present, it has close to 900k downloads on Hugging Face, which shows the support that the team received after the backlash. Another example is BharatGen’s Param-1, which, despite its small size, has only 310 downloads on AIKosh. 

The numbers notwithstanding, what happens after all these downloads? Where are the use cases?

There is still very little innovation or use case development from such models. Yes, there are developers making translation tools using Sarvam’s translation models, but nothing significant has yet been released towards ‘building AI for Bharat’.

“What to Build?”

This is not a criticism of Sarvam or BharatGen’s innovation in itself, but a testament to how Indic AI research is not attractive enough for Indian developers. 

Take BharatGen’s Project Udaan, for instance. It’s being used to translate books in schools, but the developers are struggling to innovate further because they’re unsure what else they can build on top of it. They don’t know what new use cases they can explore that haven’t already been tried by a western company, an Indian company or government initiative that could scale.

There’s an unspoken assumption that’s gaining ground here: that the current setup doesn’t encourage working on Indic languages AI. This is mostly because the return on investment is low, or in many cases, non-existent. 

Aditya Yadav, CEO of Automatski Solutions, said in a discussion, “Developers don’t give two hoots about building for Bharat because nobody in India would pay a $200/month subscription (which OpenAI charges) for their apps if they build Indic AI Apps.” 

He added that even paying ₹300/month for such AI apps is difficult. There is simply no market for this.  

A rare success story from India is Kissan AI, which also started solving problems in the West before solving for farmers in India. However, it is still in the very early stages of development, even after receiving backing from Microsoft and other investors. 

But this is just one of the rare success stories.

As Deedy Das from Menlo Ventures mentioned earlier, “There’s no real audience for this incremental work.” 

Sarvam-M wasn’t trying to win global benchmarks—it was trying to solve a fundamentally different problem: make AI accessible in Indian languages. Unfortunately, the current ecosystem rewards the former.

The Indian AI community’s obsession with MMLU, GSM-8K, and HumanEval makes one thing clear—building for Bharat doesn’t get you prestige. Unlike English-first benchmarks that dominate international conferences and unlock VC checks, Indic tasks often go unnoticed.

While the rest of the world races ahead with foundation models in English, Indian developers trying to build for native languages face a wall—one made of inadequate funding and sketchy market demand. 

On the other hand, developers are drawn to testing the models from the West to build their portfolios, which they can use to apply for jobs later. 

India’s AI ecosystem is starved for capital. 

According to AIM Research data, AI startups in the country raised just $8.2 million in Q2 2024, compared to $27 billion in the US. The government wasn’t helping either—at least not until recently. But that is also just for a selected few, not the whole developer ecosystem.

Moreover, the biggest roadblock remains the lack of Indic data for fine-tuning the models into further downstream tasks. Building a language model requires massive, clean, annotated datasets, but the same is also required for fine-tuning it for specific domains and tasks. 

It is just not available for developers to get their hands dirty with apart from just building chatbots.

What to Do?

Abhishek Upperwal, founder of Soket AI Labs, one of the startups selected by the IndiaAI Mission, told AIM that the company was generating synthetic data through translation and augmentation strategies, especially for domains like science and mathematics.

Upperwal said that the team will be able to generate 5-6 trillion unique tokens only on Indic languages, including code. In other domains, Soket expects to build a total corpus of 20 trillion tokens—a foundation large enough to train a world-class multilingual model.

This might solve the situation for Soket AI, but developers still don’t know what to build with it. While Upperwal remains optimistic about the future, he acknowledges that treating these models released by companies as academic research is more important than considering them as products. 

Most developers are left cobbling together small corpora or relying on jugaad, like translating English datasets back into Indian languages. As AIM reported earlier, “We cannot properly pursue the goal of building the next big LLM for Indic languages unless we solve the data problem.” 

And we haven’t.

Even if you manage to build a tool, the Indian market doesn’t pay. AI buyers in India are notorious for demanding endless unpaid PoCs (proofs-of-concept). It’s no surprise that startups use the term “Skip India Movement” to describe the shift toward serving only US or global clients.

Read: Free PoCs are Killing Indian AI Startups

Paras Chopra, founder of Lossfunk, admitted to banning his team from talking to Indian customers: “They’re not worth the time, effort, or resources.” This leaves even the best AI researchers stuck in the simulation mode and not building anything for users.

The reality is simple. Indic models don’t get adopted because there are no strong buyers. Even Indian IT firms are not interested in building AI for Indic languages, as their clients mostly work on English tasks.

Unlike ChatGPT or Claude, Indic AI doesn’t have paying enterprise customers; it is more of a philanthropic journey that developers need to take on by themselves. 

“The opportunity to serve somebody is the only gift that is long-lasting. Not fame, not money, not designation, not power,” Shekar Sivasubramanian, CEO of Wadhwani AI, told AIM. Sivasubramanian, who runs the non-profit, believes the goal should not be earning billions of dollars but, instead, helping somebody and making their lives better.

Most Indian firms building AI products rely on multilingual wrappers around western models. And public sector deployments are still in pilot mode. Even successful tools like CoRover’s AskDisha chatbot on IRCTC are outliers, not the norm. In fact, CoRover’s recently launched BharatGPT Mini is yet to find its exact use cases among developers.

Bhashini—India’s national language AI mission—powers 10 million translation requests a day. But that’s public infrastructure. Startups aren’t actively building on top of it because there’s very little incentive from the market—no monetisation.

“Try convincing a startup to build for Bodo,” said Bhashini CEO Amitabh Nag on AIM’s What’s the Point podcast. “That’s why the government had to step in.” 

This exposes the sad state of Indic AI developers. Very few Indian startups have the stack, let alone developers, to even try building something out. 

That’s why most of them fine-tune existing models like LLaMA or Mistral and release them as “Indian” LLMs. But when these models don’t immediately go viral, they’re criticised for not doing enough.

“There’s a whole host of Indic-language use-cases where this sovereign model would work much better,” said Pratyush Choudhury from Together Fund, adding that critics often fail to grasp how expensive and thankless this effort is.

So, when someone tries to build a foundational Indic LLM, trained on Indian data, languages, and speech, it’s treated as a side project, not a serious AI milestone. Even Sarvam’s internal team sees the problem clearly. “People critiqued the model without actually testing it,” said Harveen Singh Chadha from Sarvam AI. 

On X, others pointed out that the model answered JEE Advanced questions in Hindi correctly. Still, that didn’t stop the cynicism, which seemed fair to an extent since there was no other outcome apart from these benchmarks. 

Sovereignty Needs Builders

If we want to truly build for Bharat, VCs must fund product pilots that serve tier 2/3 India, not just English-speaking use-cases. Governments should also mandate Indic AI integration across public services and offer prize-based funding for state-level adoption.

Developers need visibility, infra credits, and real-world deployment opportunities for Indic AI work—even if it’s messy or early-stage. 

Until then, the next time an Indian team launches a multilingual LLM, don’t just ask how many downloads it got. Ask what it’s trying to solve, and whether we’ve built a system that even allows it to succeed.

Because right now, the answer is no. 

Even the startups that aim to solve Indic language tasks, do not get enough attention from investors, for which they have to develop inferior tech products. 

“We don’t want to end up building inferior tech for inclusion,” Shankar Maruwada, CEO of EkStep Foundation, told AIM. “The best technology should be for those who need it the most.” We say that India should become the AI use case capital of the world, and if that has to happen, there has to be a bigger incentive for the developers to experiment with Indic AI, which is just simply not there right now.

The post Indic AI is Not Inspiring Enough for Indian Developers appeared first on Analytics India Magazine.