Nemotron 3 Ultra Scores 30/42 at IMO 2026 in Natural Language

Nemotron 3 Ultra scored 30 out of 42 points at IMO 2026, reaching the gold-medal threshold using only natural-language proofs. No formal prover, no external tools, no internet access. The result comes from an arxiv paper by Ivan Moshkov, Stephen Ge, George Armstrong, Wei Du, Sadegh Mahdavi and Igor Gitman.

The pipeline uses three Nemotron 3 Ultra checkpoints: the general-availability model plus two post-trained specialists trained with supervised fine-tuning and reinforcement learning. The abstract describes 'an iterative search that generates, verifies, and refines candidate proofs' followed by 'a separate high-compute stage' that then selects each final submission.

Along with the paper, the authors release both post-trained checkpoints, the training data, the training and inference code, the submitted solutions, and Nemotron-IMO-Bench, described as 'a new benchmark of 200 novel olympiad-level problems.'

The abstract does not quantify the compute cost of that high-compute selection stage, nor name which problems the system missed to fall short of the 42-point ceiling. The 30-out-of-42 came from natural-language proof text scored under human IMO grading rules, not a machine-checkable verifier.