One Model Family, Two Gold-Level Results: Fine-Tuning Nemotron for IOI and IMO

What it claims

NVIDIA fine-tuned Nemotron 3 into separate specialists for competitive programming and olympiad proofs.

• IOI 2026: Nemotron-3-Ultra-CC scored 535.4/600. The run was live and prospective, under the same time, internet-access and submission limits as human contestants.
• IMO 2026: A system built from SFT and RL checkpoints of Nemotron 3 Ultra scored 30/42, above the gold threshold. It got full credit on four of the six problems, and official IMO graders marked its proofs.

The post's main argument is that neither fine-tuning nor brute-force sampling produced the medals alone. The gains came from designing the model, the data and the inference loop together.

How it works

IOI (code).

• The team curated 22,000 problems with synthetic reasoning traces.
• Nemotron-3-Nano-CC (30B total / 3B active parameters) received SFT and then RL.
• Nemotron-3-Ultra-CC (550B total / 55B active) received SFT only.
• At inference, GenCorrect runs several rounds of generate, evaluate against feedback, and refine.

IMO (proofs). Two specialists were trained from Nemotron 3 Ultra:

• SFT model: trained on 414,890 quality-filtered examples covering 15,818 unique proof problems. The data covers four tasks: proof generation, refinement, verification, and meta-verification (judging whether a proof is complete). It goes beyond final answers.
• RL model: trained on 9,597 proof problems chosen near the model's capability frontier.

The final IMO system pooled both specialists with the general-availability model. Everything ran in natural language, with no formal prover, external tools or internet access.

flowchart TD
  A[SFT checkpoint] --> G[Generate candidate proofs]
  B[RL checkpoint] --> G
  C[General model] --> G
  G --> S[Score + critique]
  S --> R[Refine best attempts]
  R --> S
  S --> F[High-compute final selection]

Evidence

IOI 2025 progression for Nano (gold threshold 438.3):

• SFT produced most of Nano's gain; RL added a small but consistent improvement.
• Ultra-CC with GenCorrect scored 502 on IOI 2025.
• One SFT epoch on Ultra was enough to beat the fully post-trained Nano on IOI, ICPC and LiveCodeBench Pro. That finding is why the IOI 2026 system used an SFT-only Ultra.

IMO development experiments:

• Both post-trained checkpoints beat the general-availability model.
• The SFT checkpoint was strongest in the first search round.
• The RL checkpoint gave the best overall single-checkpoint result.
• Combining the complementary checkpoints was worth more than drawing more samples from one checkpoint.

How to use it

IMO:

• The "Nemotron Labs IMO 2026" collection on Hugging Face contains the SFT and RL checkpoints, both training datasets, and Nemotron-IMO-Bench, a new benchmark of 200 olympiad-level problems.
• The NeMo-Skills repository has the IMO inference pipeline, prompts, the submitted proofs, and a reproducible quickstart.
• An IMO paper describes the training and the generate-verify-refine system.

IOI:

• Nemotron-3-Ultra-CC is on Hugging Face.
• The IOI evaluation and inference pipeline is in NeMo-Skills.
• An IOI paper gives the training recipe and the GenCorrect method.

The post includes no commands, licence terms or hardware requirements. Check the model cards before adopting.

Limits

• The IOI result is unofficial. It was an unsupervised benchmark and does not appear in the official IOI ranking.
• Nano-CC weights may not be published. The post names only Ultra-CC as available for code; it does not say whether Nano-CC is released.
• Compute is not quantified. The post calls training and inference "substantial", and the IMO final selection was a separate "high-compute" stage, but it gives no cost figures. Reproducing the results likely depends on large test-time budgets as well as the weights.
• Single-checkpoint results are partial. The IMO score comes from a three-model system. A single released checkpoint would not be expected to reach 30/42 alone.