Nous Research has released Nomos 1, an open-source model that scored 87 out of 120 points on the 2025 William Lowell Putnam Mathematical Competition. Based on 2024's results (the most recent published data), that score would have placed second among 3,988 participants, trailing only the top scorer who earned 90 points.
The model is a fine-tuned version of Qwen3-30B-A3B-Thinking-2507, built in collaboration with Hillclimb AI, a Y Combinator-backed startup that provides training data from IMO medalists and Putnam top performers. The 2024 Putnam had an average score of roughly 8 and a median of 2, making Nomos 1's result particularly striking for a 30-billion-parameter model.
Under identical test conditions, the base Qwen3-30B model scored just 24/120. Nous Research attributes the 3.6x improvement to training methodology rather than model scale. Nomos 1 uses a reasoning harness that runs parallel "workers" to solve problems, self-scores each attempt, then runs a finalization phase with consolidation and pairwise tournament selection to pick final answers. Solutions were graded by a human expert.
The model and reasoning harness are available now on GitHub and Hugging Face under an MIT license. Nous Research calls this its first step toward building a state-of-the-art AI mathematician.
The Bottom Line: A 30B model trained on expert math data now outscores 99.9% of undergraduate competitors on a test where the median human score is 2.
