Govt focuses on foundational AI, allocates Rs 1,500 cr, rules out tweaks to existing models

On January 31, the Centre issued a call for proposal to develop foundational models under the IndiaAI Mission

With Rs 1,500 crore earmarked under the IndiaAI Mission to support the development of indigenous foundational AI models, the Indian government on February 11 clarified that it will not consider proposals that focus solely on fine-tuning existing models.

This was conveyed to around 300 participants who attended a question and answer session on developing foundational AI models under the India AI Mission on February 11. Last month, IT minister Ashwini Vaishnaw announced that the government will be developing multiple generative AI models and deploying them within 10 months.

“Fine-tuning an existing model will not be considered,” an official said during the Q&A session with startups. Additional secretary in the IT ministry (also CEO of IndiaAI Mission) Abhishek Singh and IndiaAI Mission advisor Aakrit Vaish convened the meeting.

Officials also said that building on top of open-source models alone will not qualify as foundational model development unless significant innovation is involved.

“Depending on how ingenious you are—how do you do distilling, how do you do this—that will all have to be explained in the proposal. But just fine-tuning a LLaMA 3 or just using DeepSeek at the back end, that’s not the intent,” the official said.

The government also clarified that preference will be given to proposals that focus on multi-Indian language models. “It has to be an India-specific foundation model, and the more languages, the better,” the official quoted above said. “If we have two proposals—one addressing a single language and another covering five languages—the one with five languages will obviously be preferred.”

The government has also clarified that distillation-based approaches to model development will be assessed on a case-by-case basis.

“There are many people who say that DeepSeek is not an original LLM, it’s like they distilled it from OpenAI and other models. You could possibly do that, but you’ll need to convince the experts, the technical experts who will be evaluating it,” the official added.

The call for proposals under the mission has been kept open-ended, allowing for large language models (LLMs), multimodal models, and even smaller domain-specific models. The government has also indicated interest in large quantitative models, expanding the scope of potential projects.

On January 31, the government issued a call for proposal to build foundational models. The first cut-off for proposals is February 15, but the government has assured that there will be ongoing rounds of evaluations to accommodate more applicants.

While the funding amount per project has not been capped, officials said that teams will have to justify expenses as part of their proposals.

“Ultimately, your approach in your building... foundation models can be built within as little as six million dollars to different ranges. So depending on what you need the funding for, depending on how you are approaching the problem, what are your timelines, what is the team size… there will be like an assessment of the capabilities.”

The mission will provide compute credits to empanelled cloud vendors instead of direct hardware allocations. “India AI may fund the proposal through a direct grant or maybe through compute credits to the vendors that we have. But in your proposal, please include everything—100% of the cost, whether it’s compute for training, compute for dataset, salaries, and timelines,” the official told attendees.