The Rise of Nomos 1: A Game-Changing AI on the Putnam Math Exam

Nous Research, a San Francisco-based startup specializing in artificial intelligence, recently unveiled an open-source mathematical reasoning system known as Nomos 1. This system achieved exceptional performance on the prestigious William Lowell Putnam Mathematical Competition, a challenging undergraduate math contest. The top score in this year’s competition was 90 out of a possible 120, while Nomos 1 scored an impressive 87 points, placing it second out of nearly 4,000 participants.

Unlike larger AI models used by major tech companies, Nomos 1 boasts a compact architecture with 30 billion parameters, approximately 3 billion of which are active at any given time. The system utilizes a mixture-of-experts design based on Alibaba’s Qwen3 model.

The release of Nomos 1 signifies a significant advancement in the pursuit of AI systems capable of complex mathematical reasoning. The system’s high performance is attributed to post-training optimization and specialized reasoning techniques, rather than sheer model scale.

In a comparative analysis, it was noted that Nomos 1 significantly outperformed its base model, scoring 87 points compared to the base model’s 24 points. This underscores the importance of post-training techniques and data quality in enhancing model performance.

Validation of Nomos 1’s results was conducted through blind grading by an expert mathematician who had previously excelled in the Putnam competition. Nous Research provided anonymized submissions for grading and later published the de-anonymized files on GitHub.

The William Lowell Putnam Mathematical Competition is widely regarded as one of the most challenging mathematics competitions for undergraduate students in the United States and Canada. The exam consists of 12 questions to be solved over two 3-hour sessions, with each question worth 10 points.

Nomos 1 is a specialized version of the Qwen3 model, optimized for mathematical problem-solving and proof-writing in natural language. The system’s sophisticated reasoning harness orchestrates problem-solving strategies in two distinct phases, mirroring the structure of the Putnam competition.

The achievement of Nomos 1 highlights the trend towards smaller, more efficient AI models with advanced post-training techniques, capable of rivaling larger models developed by well-funded competitors.

In conclusion, the rapid advancements in AI-driven mathematical reasoning systems, exemplified by the performance of Nomos 1, are reshaping the landscape of mathematical problem-solving. The accessibility and efficiency of such systems hold promise for diverse applications beyond academic competitions, including formal verification, scientific modeling, and cryptographic analysis. By making these capabilities open-source, organizations can leverage cutting-edge AI mathematicians without the need for extensive computational resources.