NVIDIA's open Nemotron reaches gold-level scores at IOI and IMO

NVIDIA has described how it pushed its open Nemotron 3 models to gold-level results at the 2026 International Olympiad in Informatics and the International Mathematical Olympiad. For competitive programming, a specialist called Nemotron-3-Ultra-CC scored 535.4 out of 600 at IOI 2026, above the 361.12 gold threshold and even above the top human score of 498.27. That was an unofficial run under contestant time and submission limits, not part of the official ranking. For math, a system built on Nemotron 3 Ultra scored 30 out of 42 at IMO 2026, just past the 29-point gold line, with official IMO graders marking the proofs and full credit on four of the six problems.

The method is the part worth reading. NVIDIA says neither specialization nor extra inference compute was enough on its own; the results came from pairing a fine-tuned model with a loop that generates answers, verifies them, and refines them. The smaller Nano model, 30B total parameters with 3B active, climbed from 130 to 468 on IOI 2025 as they added supervised fine-tuning, reinforcement learning, and a test-time selection method they call GenCorrect. The math system used no formal prover, no external tools, and no internet access, relying instead on a large corpus of proof, refinement, and verification examples.

What sets this apart from a leaderboard claim is that it is open. NVIDIA put the model, both training datasets, the SFT and RL checkpoints, and a 200-problem olympiad benchmark on Hugging Face, with the IOI and IMO papers on arXiv and the inference pipelines in its NeMo-Skills repository.

Why it matters

If you work on reasoning models, the recipe is the takeaway: a tuned specialist plus a search-and-verify loop, not a single giant model doing one pass. Because the datasets and checkpoints are published, you can check the result and reuse the pieces instead of trusting a number.