Edtech firm PhysicsWallah (PW) has launched Aryabhata 1.0, an open-source small language model (SLM) built specifically to support students preparing for the mathematics section of the JEE Main exam. With 7 billion parameters, Aryabhata 1.0 has been developed to deliver focused, high-accuracy responses tailored to competitive exam formats.
Named after the ancient Indian mathematician Aryabhata, the model reflects the country’s heritage in mathematical thought and innovation. PW has made the model freely available to educators, developers, and researchers, positioning it as a resource for further innovation in education technology.
Outperforms GPT on JEE Maths
Initial benchmarks indicate impressive performance. Aryabhata 1.0 scored 90.2% on JEE Main April 2025 papers, outperforming OpenAI’s GPT-4o and GPT-4.1, which scored 43.55% and 80.44% respectively. On standard reasoning and math evaluation datasets like Math 500 and GSM8K, the model achieved 83.6% and 94.84%, respectively. Despite being smaller than large language models (LLMs), Aryabhata operates effectively within 2,000 tokens, in contrast to the 8,000-token window commonly required by larger models.
The model’s training involved 130,000 high-quality question-answer pairs drawn from PW’s internal database. It utilises a combination of supervised fine-tuning, reinforcement learning, and rejection sampling to align its outputs with the structure and logic of exam-based learning. According to PW, this targeted approach grounds the AI’s reasoning in educational practices familiar to students, potentially increasing relevance and comprehension.
Roadmap ahead
Aryabhata 1.0 is the first step in a broader roadmap. PhysicsWallah plans to release Aryabhata 2.0 later this year, expanding coverage to Physics and Chemistry and supporting curricula for JEE Advanced, NEET, and foundational academic levels.
This launch aligns with PW’s ongoing collaboration with Microsoft Research India, part of Microsoft’s $3 billion initiative to develop AI infrastructure across the country. Their joint efforts are centered on building education-optimised AI models that improve accuracy in tutoring systems and learning assistance tools.
Aryabhata’s release also comes amid growing interest in AI’s role in competitive education. A report from ByteDance highlighted how AI models such as Google’s Gemini 2.5 Pro are beginning to outperform human toppers on exams like IIT-JEE Advanced, signaling a shift in what AI can achieve in academic domains.