Nemotron Clears the Gold Medal Threshold at IMO 2026

Summary

A team fine-tuned two specialized checkpoints on Nemotron 3 Ultra using SFT and RL. They then built a pipeline (generate → verify → improve → select) that relies entirely on natural language and test-time compute. This system scored 30/42 at IMO 2026, surpassing the gold medal threshold.

Why it matters

This result is significant because it is fully open-source. The checkpoints, training data, code, submitted solutions, and a 200-question benchmark set have been released. This makes olympiad-level mathematical reasoning reproducible outside of closed labs. Furthermore, it demonstrates that iterative natural language search across multiple checkpoints can replace formal provers. This design choice lowers the competition threshold for small teams tackling difficult reasoning tasks.

Highlights

• Scored 30/42 at IMO 2026, reaching the gold medal threshold
• Fully natural language-based pipeline: no formal provers, no external tools, no internet
• Fully open-source release: two fine-tuned checkpoints, data, code, solutions, and Nemotron-IMO-Bench (200 questions)

Critical take

The phrase 'open recipe' understates the computational cost. The system relies on three checkpoints and a separate selection phase requiring high compute. Therefore, the gold medal result is as much a success of test-time scaling as it is of model quality. Additionally, losing 12 points out of 42 and the fact that the work is currently only a preprint (arXiv, awaiting DOI registration) indicate that independent reproducibility and peer review are still lacking.