NVIDIA has reported that its Nemotron 3 SI model family achieved gold-medal-level performance in both the International Olympiad in Informatics (IOI) and the International Mathematical Olympiad (IMO) for 2026. The results demonstrate that a single foundation model architecture can be specialized for distinct, high-difficulty domains using supervised fine-tuning (SFT), reinforcement learning (RL), and advanced inference techniques.

What Happened

For IOI 2026, the Nemotron-3-Ultra-CC model scored 535.4 out of 600, surpassing the gold threshold of 361.12 and the top human score of 498.27. This result was achieved during a live, prospective run under the same constraints as human contestants, including time limits and submission rules. Although unofficial and unsupervised, the run utilized a technique called GenCorrect, an iterative generate-evaluate-refine strategy. The base model, Nemotron-3-Ultra-CC, contains 550 billion total parameters and 55 billion active parameters.

The Nemotron-3-Nano-CC model (30 billion total parameters) previously reached 468 points on IOI 2025, crossing that year's gold threshold of 438.3. This progression on the 2025 benchmark informed the strategies used for the 2026 competition, where the larger Ultra model achieved the gold-medal score of 535.4.

For the IMO, NVIDIA employed a system combining SFT and RL checkpoints from Nemotron 3 Ultra. The system generated, verified, and refined proofs in natural language without using external formal provers or internet access. The final submission scored 30 out of 42, exceeding the official gold-medal threshold of 29. The SFT corpus included 414,890 quality-filtered examples across 15,818 unique proof problems, while the RL model was trained on 9,597 problems selected near the model's capability frontier. The IMO proofs were graded by official IMO graders.

Why It Matters

These results highlight a reusable recipe for specializing SI models: starting with a strong base, curating domain-specific data, applying standard post-training methods, and pairing the model with a feedback-driven inference loop. NVIDIA notes that the medals were not produced by fine-tuning or brute-force sampling alone, but by co-designing the model, data, and inference system. The company found that adaptation strategies varied by scale; for the larger Ultra model, a single epoch of SFT was sufficient to outperform the fully post-trained Nano model across several benchmarks, including IOI and LiveCodeBench Pro.

NVIDIA has released the Nemotron Labs IMO 2026 collection, which includes the SFT and RL checkpoints, training datasets, and a new benchmark called Nemotron-IMO-Bench. The Nemotron-3-Ultra-CC model is available on Hugging Face, along with papers detailing the training recipes and the GenCorrect methodology. This open-source approach allows the developer community to replicate and build upon these specialization techniques.

The Bottom Line

NVIDIA’s Nemotron 3 SI models have demonstrated world-class capability in competitive programming and mathematical proof generation through targeted fine-tuning and advanced inference workflows. The release of models, data, and benchmarks supports further research into specializing SI systems for complex, high-stakes tasks.