NVIDIA fine-tunes Nemotron 3 to gold level at both IOI and IMO 2026: a self-run 535.4 at IOI tops the best human, official graders give its proofs 30/42, and models, data and pipelines are open

Why it matters

【Key points】

On Oct 7 NVIDIA described two Nemotron 3 specialists on its Hugging Face blog. Coding: Nemotron-3-Ultra-CC (550B total, 55B active), trained on 22,000 curated problems and paired with the GenCorrect generate-evaluate-refine loop, scored 535.4/600 at IOI 2026, above the 361.12 gold line and the top human score of 498.27. Math: SFT on 414,890 proof examples plus RL on 9,597 problems, combined with the general model in a generate-verify-refine system with no formal prover, tools or internet, scored 30/42 at IMO 2026 (gold line 29), with full marks on four of six problems.

【Verifiable】

Blog: https://huggingface.co/blog/nvidia/nemotron-ioi-and-imo-2026

Model: https://huggingface.co/nvidia/NVIDIA-Nemotron-Labs-3-Competitive-Coding-550B-A55B-NVFP4

IMO proofs were graded by official IMO graders. The IOI run was NVIDIA's own, unsupervised and outside the official ranking, under contestant constraints.

【Our read】

1. The news is reproducibility: checkpoints, training data, a 200-problem Nemotron-IMO-Bench and inference pipelines are public, so olympiad gold is no longer a closed-lab badge.

2. The recipe is plain: strong base, domain data, SFT or RL, and a self-checking loop. Competition is shifting toward who builds the best verify-and-revise loop.

3. The bar is still high: 55B active parameters plus a high-compute selection stage. After OpenAI's math retractions (e/757), officially graded proofs carry more weight; the above-top-human IOI score awaits outside replication.