THE CRUNCH
NVIDIA reports that fine-tuned versions of its Nemotron model family reached gold-medal level at both the International Mathematical Olympiad (IMO) and the International Olympiad in Informatics (IOI) 2026. For IOI, a Nemotron-3-Ultra-CC specialist scored 535.4 out of 600, above the 361.12 gold threshold and the top human score of 498.27. For IMO, a system combining general, SFT and RL checkpoints in a generate-verify-refine loop scored 30 out of 42, just above the official gold threshold of 29. The IOI run was live and prospective, under the same time, internet-access and submission constraints as human contestants, though it was an unofficial benchmark outside the official ranking. The IMO proofs were graded by official IMO graders.
The two competitions test different skills: IOI demands algorithms and code that pass hidden tests under strict time and submission limits, while IMO requires rigorous natural-language proofs. For competitive programming, NVIDIA curated 22,000 problems and trained two specialists: Nemotron-3-Nano-CC, with 30 billion total and 3 billion active parameters, and Nemotron-3-Ultra-CC, with 550 billion total and 55 billion active parameters.
The company's IOI 2025 experiments show what the specialisation bought. Nano improved from 130 points before post-training to 280 after supervised fine-tuning and 291 after reinforcement learning, then reached 468 with GenCorrect, an iterative generate-evaluate-refine strategy, crossing the 438.3 gold threshold. Adaptation also differed by scale: one SFT epoch was enough for the stronger Ultra model to beat the fully post-trained Nano across IOI, ICPC and LiveCodeBench Pro.
For IMO, the SFT corpus contained 414,890 quality-filtered examples across 15,818 unique proof problems, covering proof generation, refinement, verification and meta-verification. The RL model trained on 9,597 problems near the model's capability frontier. The SFT checkpoint was strongest in the first search round while the RL checkpoint achieved the best overall single-checkpoint result, so the final system used both alongside the general model. NVIDIA has released the models, data and recipes on Hugging Face.