While India doesn’t yet have a frontier AI model, Indian startups are slowly developing smaller models that can work well on specific use-cases.
Indian ed-tech startup PhysicsWallah has helped in the creation of Aryabhata 1.0, an AI model that’s designed to solve IIT JEE mains questions. On the IIT JEE paper that it was tested on, Aryabhata 1.0 scored 90.2 percent marks, ahead of older American and Chinese models. The model was trained by AthenaAgent, a company which performs post-training on production-grade AI agents.
Aryabhata 1.0 is a 7B parameter compact math reasoning model built specifically for Indian exams. It was trained on a single H100 GPU, using custom RLVR (Reinforcement Learning with Verifiable Rewards) training on 130K+ curated JEE problems. The model has a context window of 4k tokens. Aryabhata 1.0 claims to use fewer tokens than many other models to be able to solve the JEE problems
While the results are impressive, they aren’t close to the results achieved by frontier models. Top models like Gemini Pro 2.5 and OpenAI’s o3 can not only solve IIT-JEE Mains problems, but Gemini Pro 2.5 had managed a score of 332/360 on JEE Advanced, which would’ve given it the first rank in the prestigious exam. While Aryabhata 1.0 currently is able to work on only JEE Mains, plenty of other models from US and Chinese labs can not only do well on JEE mains, but also get high marks on the JEE Advanced.
But these are some good baby steps for India’s startup ecosystem towards embracing AI. PhysicsWallah teaches millions of lakhs of students every year, and an AI model that can solve IIT JEE problems could come in handy in a variety of ways. Recently, Zoho too had also announced similar AI agents that worked on specific use-cases. Meanwhile, companies like Sarvam are looking to train bigger models and are receiving compute resources from the government. And while it might take Indian companies several years to catch up to where the US and China are at the moment, it appears that they’re taking the first steps towards getting there.