After much debate on India’s own foundational model, the country’s start-ups seem keen to take up the mantle of building their own AI LLMs, while also leveraging the information made public by companies like DeepSeek. Start-ups such as CoRover AI, Sarvam AI, Krutrim AI, etc. are looking to build an AI model from scratch
CoRover AI, owner of the indigenous BharatGPT chatbot, said it is taking a flexible and use-case driven approach wherein it plans to leverage existing LLM models where applicable, while also investing in its sovereign AI LLM.
“Both approaches are necessary. For frontier models targeting specific industries or use cases, fine-tuning a base model with relevant data enables faster and more cost-effective implementation. Conversely, for broader applications with abundant, high-quality data, building from scratch is preferable to mitigate potential data theft litigation risks, which the popular LLM companies are currently facing,” said Ankush Sabharwal, Founder and CEO of CoRoverAI. Businessline also reached out to Sarvam AI for their views, but the company declined to comment.
Kiran Chandra, Founder of open source association Swechha and Centre Head at Viswam.AI centre of excellence, said recent developments by DeepSeek, particularly, have enabled researchers and developers in building foundational models by allowing free re-use and modification of the DeepSeek model under the MIT licence. Unlike restrictive licences, the MIT licence allows users to build better models on top of DeepSeek’s foundation. However, DeepSeek stands out in comparison to other open source models in that it also shares its model weights.
“DeepSeek’s data sets are not publicly available, but the weights of the model allows us to study the model. We can also leverage the DeepSeek algorithm. This means we can basically build our own foundational model. For India, this presents a unique opportunity. By focusing on creating high-quality, digital-first datasets tailored to local contexts, India can leverage this open-weight model to develop world-class AI systems,” said Chandra.
In terms of data sets, Chandra said India should focus on using its own data sets that can inform the AI model about the country’s unique cultural contexts.
“One of the reasons LLMs hallucinate is because they don’t understand the cultural context. We want to fix that through digital first data collection. If the LLM is only going to serve copy-editing, it’s going to be a glorified search engine/ copyedited tool. But when you consider the use-case for a farmer, a domestic housewife, some amount of context and culture needs to be embedded. We want to solve problems for resource poor languages and so, we are taking up this approach to building the foundational model,” said Chandra.
On the other hand, Ganesh Katrapati, Co-Founder of Alonzo AI services and consulting firm, was unconvinced that building foundation models from scratch could add any value for start-ups or to their customers at this stage.
“There is a lot of pressure on the Indian government and start-ups to do what DeepSeek has done, but I don’t find value in replicating this kind of model. This is turning into a huge PR exercise into building so-called Indian models for Indian languages. They’re only thinking about adapting to Indian languages. One should evaluate based on research whether the existing models can be fine-tuned to perform well with Indian languages. If there’s a data problem, you can just add more data and then it’ll work better. I don’t think one should spend money in duplicating either DeepSeek or Open AI.
He added that the government should also focus on funding in fundamental research, rather than focusing on Indic LLMs, as that would help start-ups more financially.
On fiscal hurdles for start-ups, Chandra said that the government should bring down the compute cost and consider releasing the data sets.
Similarly, when asked about concerns regarding the cost of building foundational AI models, Sabharwal said the cost is “genuinely” much lower compared to earlier expenses, allowing start-ups to leverage existing advancements while eliminating the need for ground-up research and development. Further, he said that a homegrown LLM will empower Indian start-ups and researchers to innovate, driving growth and job creation. He also stressed the need for India to develop its own foundational LLMs for Indic languages.
“With 22 official languages and numerous dialects, India’s linguistic diversity is unparalleled. However, global LLMs primarily cater to Western languages, leaving a significant gap in addressing India’s unique language needs,” said Sabharwal.
Up until December 2024, Indian companies were critical about the idea of building their own foundational models. Even K Krithisavasan, CEO of Tata Consultancy Services, had said there’s no huge advantange for India in going down this road. At the time, this seemed to be the general sentiment in the Indian ecosystem. However, following Chinese company DeepSeek’s demonstration of AI affordable solutions, start-ups at least, seem to have had a change of heart.
