NVIDIA Nemotron achieves gold-level results at IOI and IMO, same model family tops both competitions

On October 7, 2026, the NVIDIA team published an article titled "One Model Family, Two Gold-Level Results: Fine-Tuning Nemotron for IOI and IMO" on the Hugging Face blog, authored by Aleksander, Igor Gitman, Sean Narenthiran, Mehrzad Samadi, and Somshubra Majumdar. The article notes that the International Olympiad in Informatics (IOI) tests algorithmic and code capability under strict time and submission constraints with hidden tests, while the International Mathematical Olympiad (IMO) demands rigorous natural-language proofs; achieving gold in both means a model has broader general capability.

In specific results, at IOI 2026 the team used Nemotron-3-Ultra-CC with the GenCorrect reasoning loop, under the same time, internet access, and submission limits as human contestants, achieving 535.4/600, surpassing both the 361.12 gold threshold and the 498.27 top human score. However, this result is an unofficial, unsupervised benchmark, not counted in IOI's official rankings. At IMO 2026, the team combined Nemotron 3 Ultra's general model with SFT and RL checkpoints into a generate-verify-refine system, achieving 30/42, exceeding the 29-point official gold threshold, with 4 of 6 problems at full marks, and submitted proofs graded by official IMO graders. The system ran entirely in natural-language environment, without formal provers, external tools, or internet access.

The article emphasizes that both results follow a reusable specialization recipe with four steps: select a strong Nemotron base model; curate domain-specific problems and high-quality reasoning samples; apply standard post-training methods like SFT and RL; combine the expert model with reasoning loops that generate, evaluate, and refine candidate answers. For competitive programming, the team curated 22,000 problems with synthetic reasoning samples and trained two expert models: Nemotron-3-Nano-CC (30B total, 3B activated parameters) received SFT and RL; Nemotron-3-Ultra-CC (550B total, 55B activated parameters) received SFT only.

The article shows IOI 2025 progression: Nano improved from 130 (pre-finetune) to 280 after SFT, 291 after RL, and 468 with GenCorrect, exceeding the 438.3 gold threshold; Ultra-CC reached 502 with the same test-time strategy. The research also found that adaptation methods can differ by model size: SFT gives Nano its main gains, while the stronger Ultra model needs only one SFT training cycle to surpass fully post-trained Nano on IOI, ICPC, and LiveCodeBench Pro, directly informing the Ultra-CC system for IOI 2026.

For IMO, the SFT corpus contains 414,890 high-quality samples filtered from 15,818 unique proof problems, covering proof generation, improvement, validation, and meta-validation, teaching the model to construct arguments, identify gaps, respond to criticism, and judge proof completeness; the RL model is trained on 9,597 proof problems near the capability frontier. The SFT checkpoint is strongest in first-round search, the RL checkpoint achieves the best single-shot performance, and the two complement each other, with the final system using both experts plus the general model.

The article argues that specialization and test-time compute are synergistic: GenCorrect converts gains from finetuning into larger improvements through multi-round feedback, and combining complementary SFT and RL checkpoints in IMO beats simply drawing more samples from a single checkpoint. Related resources include the Nemotron Labs IMO 2026 collection on Hugging Face (with SFT and RL checkpoints, two training datasets, and the new Nemotron-IMO-Bench evaluation of 200 problems), the Nemotron-3-Ultra-CC model, the IOI and IMO papers, and the NeMo-Skills code repository with reasoning pipelines, prompts, submitted proofs, and reproducible quickstart examples.