NVIDIA Nemotron Olympiad Recipe: What the Open Release Means

NVIDIA published a detailed write-up on October 7, 2026 explaining how its Nemotron 3 models reached gold-medal territory at two of the hardest academic competitions in the world, the International Olympiad in Informatics (IOI) and the International Mathematical Olympiad (IMO). The bigger news for developers is what came with it: NVIDIA open-sourced the models, the training data and the inference pipelines behind both results on Hugging Face, so anyone can study or build on the recipe.

• IOI 2026: Nemotron-3-Ultra-CC scored 535.4 out of 600, above the gold line of 361.12 and the top human score of 498.27, according to NVIDIA.
• IMO 2026: Nemotron 3 Ultra scored 30 out of 42, one point above the gold threshold of 29, with full credit on 4 of 6 problems.
• Models: Nemotron-3-Ultra-CC has 550B total and 55B active parameters; Nemotron-3-Nano-CC has 30B total and 3B active.
• Open release: the coding model, IMO checkpoints and datasets, a 200-problem benchmark called Nemotron-IMO-Bench, and the NeMo-Skills pipelines.

What Did Nemotron Achieve at IOI and IMO 2026?

On the programming side, NVIDIA says its competitive-coding variant, Nemotron-3-Ultra-CC, combined supervised fine-tuning with a technique it calls GenCorrect to score 535.4 out of 600 on the IOI 2026 problem set. NVIDIA is careful to note that this was a live, prospective run, but an unofficial and unsupervised one that was not part of the official IOI rankings. That honesty matters, and the comparison to the top human score should be read with that context.

The maths result used the general Nemotron 3 Ultra model inside a generate, verify and refine loop that draws on both supervised and reinforcement-learning checkpoints. It scored 30 of 42 points, clearing the gold line of 29 and earning full marks on four of the six problems.

How Was the Open Recipe Built?

The most useful line in the post is NVIDIA's point that the medals did not come from fine-tuning alone. They came from designing the model, the data and the inference loop together. The training data sizes show how much curation went in:

• IOI: about 22,000 curated programming problems with synthetic reasoning traces.
• IMO supervised fine-tuning: 414,890 quality-filtered examples drawn from 15,818 unique problems.
• IMO reinforcement learning: 9,597 proof problems.

Both models use a mixture-of-experts design, so only a fraction of the parameters run for each token. That is the same efficiency trick behind most recent open frontier models, including the Nemotron 3 Ultra 550B launch we covered in June.

Why Does an Open Olympiad Recipe Matter?

Olympiad-level reasoning results have mostly arrived as closed demonstrations. Here, researchers get the checkpoints, the curated datasets and the exact inference pipeline, plus Nemotron-IMO-Bench, a fresh set of 200 olympiad-level problems for testing future models without worrying that the answers leaked into training data.

That opens the door for universities and smaller labs to reproduce, critique and improve the approach rather than take it on faith. It also fits a broader trend of AI maths work becoming more verifiable, following efforts like AI-checked Lean proofs. For more on open models, see our AI coverage.

What Should Developers Try First?

If you want to experiment, the NeMo-Skills repository is the practical starting point, since it includes the IMO inference pipeline, prompts, the submitted proofs and the IOI evaluation setup. The Nano-CC model, with just 3B active parameters, is the realistic entry point for teams without a large GPU cluster. All scores above come from NVIDIA and have not yet been independently reproduced.

Sources: Hugging Face blog (NVIDIA) — October 7, 2026; Tech Times — September 5, 2026.