2026-10-07 — views
Nemotron Models Achieve Gold Medal Performance in IOI and IMO
Read this because This demonstrates that specialized fine-tuning combined with iterative inference loops can push open-weight models to elite competitive levels without external tools.
NVIDIA fine-tunes Nemotron models to reach gold-level results in both the International Olympiad in Informatics and International Mathematical Olympiad.
Nemotron Achieves Dual Gold-Level Results
NVIDIA researchers have demonstrated that the Nemotron model family can achieve gold-medal level performance in two distinct, high-difficulty competitions: the International Olympiad in Informatics (IOI) and the International Mathematical Olympiad (IMO). Published on October 7, 2026, the results highlight the adaptability of Nemotron 3 as a foundation for building world-class specialist models through supervised fine-tuning (SFT), reinforcement learning (RL), and feedback-driven inference.
IOI 2026 Performance
For the IOI 2026, the team utilized Nemotron-3-Ultra-CC, a model with 550 billion total parameters and 55 billion active parameters. This model underwent supervised fine-tuning and was paired with a technique called GenCorrect, an iterative generate-evaluate-refine strategy. The system achieved a score of 535.4 out of 600, surpassing the gold threshold of 361.12 and the top human score of 498.27. This result came from a live, prospective run under the same time, internet-access, and submission constraints as human contestants, though it remains an unofficial benchmark. Earlier experiments with the smaller Nemotron-3-Nano-CC model showed that SFT and RL significantly improved scores, with GenCorrect pushing the Nano model past the gold threshold as well.
IMO 2026 Performance
The IMO 2026 project focused on rigorous natural-language proofs. Starting from Nemotron 3 Ultra, the team trained specialists using SFT and RL. The SFT corpus included 414,890 quality-filtered examples across 15,818 unique proof problems, teaching the model to construct arguments, identify gaps, and judge completeness. The RL checkpoint was trained on 9,597 proof problems near the model’s capability frontier. The final system combined these checkpoints with the general model to generate, score, critique, and refine candidate proofs. Operating entirely in natural language without formal provers or external tools, the system scored 30 out of 42 points, exceeding the official gold threshold of 29 and receiving full credit on four of the six problems. These proofs were graded by official IMO graders.
A Reusable Specialization Recipe
The authors outline a four-part recipe for specialization: start with a strong base model, curate domain-specific problems and reasoning traces, apply post-training methods like SFT and RL, and pair the model with an inference loop that generates, evaluates, and improves answers. This approach proves that adaptation does not require building new foundation models for every challenge. Instead, Nemotron can be specialized for specific tasks.
Practitioner note
Developers should explore the Nemotron Labs IMO 2026 collection and the NeMo-Skills repository on Hugging Face to access the training datasets, checkpoints, and reproducible inference pipelines. By implementing similar generate-verify-refine loops with specialized checkpoints, engineers can potentially replicate these high-performance results in other complex domains requiring rigorous reasoning.