NVIDIA says Nemotron specialists reached gold level in maths and coding Olympiads

NVIDIA says it fine-tuned its Nemotron models into specialist systems that reached gold-medal level at both the 2026 International Mathematical Olympiad and International Olympiad in Informatics. The results point to a practical recipe: adapt a capable model for a demanding subject, then give it a system for generating, checking and improving answers.

NVIDIA Watch analysis

What happened

In a Hugging Face post, NVIDIA describes separate projects for mathematical proofs and competitive programming. The company says both began with Nemotron 3 and used supervised fine-tuning; the maths project also trained a reinforcement-learning specialist, while the coding work used an iterative generate-evaluate-refine approach called GenCorrect.

The IMO system scored 30 out of 42, including full marks on four of six problems, exceeding the official gold threshold. Its proofs were graded by official IMO graders. At the IOI, NVIDIA says its system scored 535.4 out of 600 in a live run under the competition’s time, internet-access and submission constraints. That result was unofficial and did not count towards the official ranking.

Key findings

• Fine-tuning lifted the coding score
  NVIDIA says its smaller specialist rose from 130 to 280 points after supervised fine-tuning, then to 291 after reinforcement learning.
• Inference-time feedback added a larger jump
  With GenCorrect, the smaller model reached 468 points on IOI 2025 problems, above the 438.3-point gold threshold; a larger specialist reached 502.
• The maths system combined complementary specialists
  NVIDIA says its final system used supervised-fine-tuned and reinforcement-learning models alongside the general model to generate, critique and refine proof candidates.

Why it matters

The results make a useful distinction between making a model better and building a system that can use it well. NVIDIA’s account suggests that targeted training supplies stronger candidates, while repeated generation, checking and refinement can turn those candidates into much better competition results. For teams adapting models to specialised work, the inference loop may matter as much as the fine-tuning recipe.

There is also unusually tangible material for others to examine: NVIDIA says it has released the IMO checkpoints and datasets, a 200-problem benchmark, and IOI training and inference materials. The company’s results are not a guarantee that the same approach will transfer to everyday coding or mathematics, but they give researchers something more useful than a medal claim alone: a recipe and artefacts to test.

Our read

This is a strong demonstration of model specialisation working hand in hand with test-time search. The impressive part is not simply that a model reached a medal threshold, but that the method combines familiar training techniques with a system that checks and improves its own attempts. The IOI result remains unofficial, so keep that distinction firmly in view; the IMO proofs, meanwhile, were graded by official markers.

If you work on specialised models, the published datasets, checkpoints and pipelines are the practical place to start. The next question is whether outside teams can reproduce the gains without needing competition-scale compute.

What to watch

• Whether independent teams reproduce the IMO and IOI results using the released materials.
• How the approach performs on new problems, rather than competition sets used in development.
• Whether the same generate-check-refine methods transfer to less neatly scored real-world tasks.

Discussion spark: For specialised AI systems, which deserves more credit for the result: fine-tuning the model, or the generate-check-refine loop wrapped around it?

Independent WittyWires tracker for public updates about NVIDIA. Not affiliated with or endorsed by NVIDIA; this is not an official account.