Fractal Analytics Makes a Bid for the First Large Reasoning Model in India

At a total project cost of INR 118.8 crore, Fractal Analytics, an artificial intelligence startup based in Mumbai, has planned to construct India's first large reasoning model (LRM). Fractal has proposed to the IndiaAI Mission that the government provide INR 76.6 crore in external finance for the project. A small model (2–7 billion parameters), a medium model (20–32 billion parameters), and a big state-of-the-art model (70 billion parameters) with up to 1 trillion training tokens will all be part of the planned LRM series. Srikanth Velamakanni, the inventor of Fractal, confirmed the news to a media outlet, saying that India should prioritise post-training models that are capable of "thinking and reasoning" as opposed to pre-trained models.

What are LRMs?

Large language models (LLMs) that specialise in advanced thinking, problem-solving, and decision-making activities are known as LRMs. Compared to OpenAI's o1 and o3 reasoning models, Fractal's proposed model is far larger in terms of size. DeepSeek R1 is now the largest reasoning model in the world, with 671 billion parameters. India might lead the world in AGI (artificial general intelligence) and catch up to nations like the US and China with the help of the planned LRM. The days of pre-training are finished, according to Velamakanni. Building systems that can operate with pre-trained models and complete challenging real-world tasks through improved planning and reasoning is currently the main focus of the race for stronger AI. According to Velamakanni, Fractal wants to incorporate Indian LLMs and create reasoning skills from scratch on open-source LLMs with permissive licences.

IndiaAI Mission

During the second round of bidding, which ended on March 15, the IndiaAI Mission received 120 applications to create sovereign AI foundation models. 187 applications have been submitted thus far, with 67 ideas having been received in the first round, as per various media reports. A senior government official told a prominent media outlet that the Ministry of Electronics and Information Technology (MeitY) hopes to approve at least a handful of the applications by the end of the month. In order to develop a world-class reasoning model in STEM, coding, medicine, and agentic systems, Fractal also intends to produce and make available a variety of datasets from top tests in India, including JEE Advanced, NEET-PG, National Olympiads, CAT, GATE, and others. Due to data saturation and performance plateaus, scaling train-time computation presents diminishing returns. Thus, according to the Mumbai-based advanced research firm, increasing inference-time computation is a cost-effective strategy and creates a quicker route to AGI for India.