One Model Family, Two Gold-Level Results: Fine-Tuning Nemotron for IOI and IMO
- Source
- huggingface.co
- Date

Context
NVIDIA reports that fine-tuned versions of its Nemotron 3 reached gold-level results at two 2026 contests, in its Hugging Face post. On the International Olympiad in Informatics (IOI), a programming contest, a live but unofficial run scored 535.4 of 600 against a 361.12 gold threshold. On the International Mathematical Olympiad (IMO), the system wrote natural-language proofs, graded by official IMO graders according to the post, and scored 30 of 42 against a gold threshold of 29.
The post credits specialized supervised and reinforcement-learning together with a run-time loop that generates candidate solutions, checks them and refines them. For math, a separate high-compute stage chose the final submission, so the result reflects heavy effort. The IOI run sat outside the official ranking. NVIDIA's open recipe for IMO gold covers the math side in more detail.
- Model family — A group of related models developed under a shared name or approach.
- Training — The process that adjusts a model using examples and a measure of how well it performs.
- Inference — Using a trained model to produce an output from a supplied input.
- test-time compute — Spending more computation when the model answers — thinking longer, trying multiple attempts — to buy accuracy without training a bigger model.
articleAn Open Recipe for IMO Gold: Training Nemotron for Olympiad MathematicsIvan Moshkov, Stephen Ge, George Armstrong, Wei Du, Sadegh Mahdavi, Igor Gitman
articleOptimizing Agent Reliability and Tool Calling Efficiency for Chip Design Using NVIDIA NeMo GymTanya Lenz
articlePost-Training NVIDIA Nemotron to Increase Accuracy in Your Scientific DomainGautham Dharuman
Checking sign-in…
Loading comments…