TLDR: Ed-tech firm PhysicsWallah has launched Aryabhata 1.0, an AI model specifically designed for Indian competitive exams, which scored an impressive 90.2% in the April session of the IIT JEE Mains mathematics paper, marking a significant step for AI in Indian education.
PhysicsWallah (PW), a prominent Indian ed-tech startup, has announced the development and successful performance of its new artificial intelligence model, Aryabhata 1.0. This compact math reasoning model achieved a remarkable 90.2% score on the IIT JEE Mains examination’s April session, following an 86% score in the January 2025 session. The model’s creation was supported by AthenaAgent, a company specializing in post-training production-grade AI agents.
Aryabhata 1.0 is a 7-billion parameter causal decoder-based model, specifically optimized for high-stakes Indian competitive exams like JEE Mains. Its development underscores a growing trend among Indian startups to create specialized AI models for particular use-cases. The model was trained on a single NVIDIA H100 GPU, utilizing a custom Reinforcement Learning with Verifiable Rewards (RLVR) approach on over 130,000 curated JEE problems. The training pipeline also incorporated techniques such as Model Merging, Rejection Sampling, and Supervised Fine-Tuning (SFT) to enhance its reasoning capabilities and align with pedagogical patterns for competitive exams.
Prateek Maheshwari, co-founder of PW and chair of the Indian Edtech Consortium, emphasized the potential of AI in education, stating, “Sometimes, the right support at the right moment can change a student’s entire path. At PW, we believe AI can offer that support, if it’s built with care, context, and purpose.” He also indicated that Aryabhata 1.0 is part of a broader strategic roadmap, with plans to expand its capabilities to cover JEE Advanced and a wider array of mathematical domains. PhysicsWallah is actively inviting educators, developers, and researchers to test the model and provide feedback for its refinement, aiming to build a comprehensive suite of subject-centric Small Language Models (SLMs).
Also Read:
- OpenAI’s Experimental AI Achieves Gold Medal Performance at International Mathematical Olympiad
- IIT Patna Unveils Six-Month Professional Certificate Program in Generative AI
While Aryabhata 1.0’s performance is notable, especially given its compact size and efficient training (operating effectively around a ~2K token window compared to ~8K required by other reasoning models), it is acknowledged that it does not yet match the capabilities of frontier AI models like Google’s Gemini Pro 2.5 or OpenAI’s o3. These advanced models have demonstrated superior performance, with Gemini Pro 2.5, for instance, scoring 332 out of 360 on JEE Advanced, a score that would have secured the top rank in the prestigious exam. Nevertheless, Aryabhata 1.0 represents a significant ‘baby step’ for India’s burgeoning startup ecosystem in its embrace of AI for specialized applications, following similar initiatives by companies like Zoho in developing use-case specific AI agents.
