Nous Research Just Released Nomos 1, an Open-Source AI that Ranks Second on the Notoriously Brutal Putnam Math Exam

Why It Matters

Efficient, specialized models can match trillion‑parameter rivals, giving enterprises advanced math reasoning without massive compute costs. This democratization could speed verification, scientific modeling, and cryptographic analysis across industries.

Key Takeaways

• Nomos 1 scored 87/120, ranking #2 in Putnam
• Model uses 30B parameters, only 3B active via MoE
• Specialized post‑training boosted score from 24 to 87
• Runs on consumer hardware, unlike trillion‑parameter rivals
• Open‑source harness enables reproducible math reasoning

Pulse Analysis

The Putnam Competition has long been the gold standard for testing deep mathematical insight, with a median score of just two points out of 120. Nomos 1’s 87‑point performance not only eclipses the vast majority of human participants but also narrows the gap with frontier models such as DeepSeekMath‑V2 and Google’s Gemini, which achieve higher raw scores but demand massive compute clusters. By delivering near‑elite results on a publicly available benchmark, Nous Research signals that the race for AI mathematicians is no longer limited to hyperscale labs.

Technically, Nomos 1 leverages a 30‑billion‑parameter mixture‑of‑experts architecture derived from Alibaba’s Qwen3, activating roughly three billion parameters per inference. The breakthrough stems from intensive post‑training techniques and a two‑phase reasoning harness: a solving phase that self‑critiques submissions and a finalization phase that consolidates and selects the best proof. This design transforms a modest base model that scored only 24 points into a system capable of eight perfect solutions, all while running on consumer‑grade GPUs, dramatically lowering the entry barrier for sophisticated mathematical AI.

For businesses, the open‑source release under Apache 2.0 means they can integrate state‑of‑the‑art mathematical reasoning directly into internal pipelines without relying on expensive cloud APIs. Applications range from formal verification of software and hardware designs to automated theorem proving in research and cryptographic analysis. As more efficient models like Nomos 1 prove their worth, the industry can expect a surge in specialized AI tools that combine high performance with affordable deployment, reshaping how organizations approach complex problem solving.