Nvidia Nemotron Wins Gold at IMO and IOI 2026 — Open Release

Nvidia has announced that its Nemotron 3 model family achieved gold-medal-level results at two of the most demanding intellectual competitions in the world: the International Mathematical Olympiad (IMO) and the International Olympiad in Informatics (IOI), both in 2026. The announcement, published on Hugging Face, is notable not just for the results but for what comes with them — a complete, open-weight release of every component used to get there.

The IOI system, Nemotron-3-Ultra-CC, scored 535.4 out of 600 under the same time limits, internet restrictions and submission constraints that human contestants faced. The gold threshold was 361.12; the top human score was 498.27. At IMO 2026, a pipeline combining the base Nemotron 3 Ultra model with two specialist fine-tunes scored 30 out of 42 points — above the official gold threshold of 29 — with proofs graded by official IMO judges. It is worth being precise here: neither result came from a single model acting alone. The medal-level performance in both competitions emerged from an ensemble approach pairing fine-tuned specialist models with iterative inference loops that generate, critique and refine candidate answers.

“Easy to fine-tune should mean more than making a checkpoint trainable. It should mean that a capable foundation model can be adapted to a demanding domain with a clear, reusable recipe,” wrote Nvidia.

That recipe, as Nvidia describes it, has four steps: start with a strong Nemotron base model, curate domain-specific problems and high-quality reasoning traces, apply supervised fine-tuning (SFT) and — where beneficial — reinforcement learning (RL), then pair the resulting specialist with an inference loop that searches, verifies and improves. The team did not build new foundation models from scratch for each competition. They specialized what already existed.

What the Training Actually Looked Like

For competitive programming, Nvidia trained two models on a corpus of 22,000 curated problems. Nemotron-3-Nano-CC, a sparse model with 30 billion total parameters but only 3 billion active, received both SFT and RL. Nemotron-3-Ultra-CC, at 550 billion total and 55 billion active parameters, received SFT only. The development results on IOI 2025 benchmarks trace a clear arc: Nano went from 130 points before any post-training, to 280 after SFT, to 291 after RL, to 468 using GenCorrect — Nvidia’s iterative generate-evaluate-refine strategy — well above the 438.3 gold threshold. Ultra-CC, with the same inference strategy, reached 502. The competition system scored 535.4.

The math side was more complex. The IMO SFT corpus contained 414,890 quality-filtered examples drawn from 15,818 unique proof problems, designed to teach not just final answers but the full arc of mathematical argument: generating proofs, identifying gaps, responding to critiques and verifying whether a proof is actually complete. A separate RL model trained on 9,597 problems chosen near the model’s capability frontier. The two checkpoints had complementary strengths — the SFT model performed best in early search rounds; the RL model achieved the best single-checkpoint result overall — so the final competition system used both alongside the general Nemotron 3 Ultra model. The entire pipeline operated in natural language, with no formal proof assistant, no external tools and no internet access.

Where This Fits in the AI Olympiad Race

The contest between AI systems and human olympiad competitors has become one of the most closely watched benchmarks at frontier AI labs. OpenAI reported gold-medal-level IOI performance with its o1-ioi system in 2024, a dedicated fine-tune that also relied on external tools. Several proprietary models have since claimed comparable results, often without disclosing their methods. The current IOI leaderboard includes systems from Google and OpenAI solving problems across multiple competition years at very high rates.

What makes the Nemotron release stand apart is not that it tops every leaderboard — it does not. What distinguishes it is reproducibility. Nvidia has released the SFT and RL checkpoints, both training datasets, the GenCorrect inference pipeline, submitted competition proofs, evaluation code and Nemotron-IMO-Bench, a new benchmark of 200 novel olympiad-level problems. The IOI result was unofficial and not included in the official ranking, but it was conducted under real competition conditions without special access. The IMO proofs were judged by the actual competition graders.

Why This Matters If You’re a Student or Early-Career Researcher

This release is unusually actionable. The base Nemotron 3 Ultra model is available on Hugging Face under a commercially permissive license and accessible via API on platforms like OpenRouter. But the more valuable assets may be the artifacts surrounding it. Graduate students and self-taught researchers now have access to a documented, end-to-end post-training pipeline — from raw problem curation through SFT, RL and inference-time search — that produced verifiably gold-medal results. That is a live case study in the techniques currently most in demand at AI labs: SFT versus RL tradeoffs, test-time compute scaling and generate-verify-refine inference design.

For competitive programmers and math olympiad participants, Nemotron-IMO-Bench offers 200 new hard problems to train and evaluate against — a resource that did not exist a week ago. For anyone building AI-adjacent skills, the NeMo-Skills repository includes the full inference pipeline, prompts and a reproducible quickstart. Nvidia’s central argument is that you do not need to be a frontier lab to replicate frontier-adjacent results: a strong open base model, carefully curated data and a well-designed inference loop may be enough.

The takeaway for students is straightforward. The gold medals were not produced by raw model scale alone, and they were not produced by clever prompting alone. They came from deliberately co-designing the model, the training data and the inference process — a workflow that is now documented, open and reproducible. That is a template worth studying.

Source: Hugging Face