Fractal Analytics pitches for India’s first large reasoning model

Fractal Analytics, a Mumbai-based artificial intelligence company, has proposed to build India’s first large reasoning model (LRM) at a total project cost of Rs 118.8 crore.

In a proposal to the IndiaAI Mission, Fractal has sought an external funding of Rs 76.6 crore from the government for the initiative.

The proposed LRM series will include a small model (2-7 billion parameters), medium model (20-32 billion parameters) and a large state-of-the-art model (70 billion parameters) with up to 1 trillion training tokens.

Confirming the development to ET, Fractal founder Srikanth Velamakanni said India needs to focus on post-training models that can “think and reason” rather than on pre-trained models.

LRMs are a specialised class of large language models (LLMs) focused on performing advanced reasoning, problem-solving and decision-making tasks.

Live Events

In terms of scale, Fractal’s planned model is significantly larger than OpenAI’s o1 and o3 reasoning models. Currently, the largest reasoning model globally is DeepSeek R1, with 671 billion parameters.

The proposed LRM may help India catch up with countries like the US and China, and lead the world towards AGI (artificial general intelligence).

“The era of pre-training is over,” Velamakanni said. “The race for better AI is now focused on building systems that can work with pre-trained models and accomplish complex real-world tasks by better planning and reasoning.”

Fractal aims to develop reasoning capabilities from scratch on open source LLMs with permissive