NVIDIA explains Nemotron’s olympiad-level training recipe

What is changing?

NVIDIA’s October 7 technical post describes coding and mathematics systems built from Nemotron 3. It reports 535.4/600 in an IOI 2026 run and 30/42 in an IMO 2026 evaluation, above the gold thresholds of 361.12 and 29.

The IOI run was unofficial and unsupervised, outside the official ranking; this does not mean an official medal was won. Official graders assessed the IMO proofs. Results combine fine-tuning, reinforcement learning and an inference workflow that generates, checks and improves candidates.

Model weights, datasets and NeMo-Skills recipes are available through the linked repositories. Today’s news explains the method; it does not claim every file was first released today. AI Radar perspective: The useful lesson for specialist agents is to pair generation with verification and revision.