Nemotron 3: Gold-Level Performance in IOI and IMO 2026

NVIDIA has achieved gold-medal level performance in both the 2026 International Olympiad in Informatics (IOI) and the 2026 International Mathematical Olympiad (IMO) by specializing the Nemotron 3 model family. This success demonstrates that a strong foundation model can be adapted into high-performing domain specialists through a combination of supervised fine-tuning (SFT), reinforcement learning (RL), and iterative inference loops.

Performance Results for IOI and IMO 2026

The Nemotron-based systems achieved scores that exceeded the official gold-medal thresholds for both competitions:

Competition Nemotron Specialization Result
IOI 2026 Nemotron-3-Ultra-CC (SFT + GenCorrect) 535.4/600 (Gold threshold: 361.12; Top human score: 498.27)
IMO 2026 Nemotron 3 Ultra (General, SFT, and RL checkpoints in generate-verify-refine system) 30/42 (Gold threshold: 29)

For IOI 2026, the results were obtained from a live, prospective run under the same constraints as human contestants (time, internet access, and submission limits), though it was an unofficial, unsupervised benchmark. For IMO 2026, the proofs were graded by official IMO graders.

The Specialization Recipe

NVIDIA utilizes a four-part reusable recipe to adapt Nemotron base models into demanding domain specialists:

  1. Foundation: Start with a strong Nemotron base model.
  2. Data Curation: Curate domain-specific problems and high-quality reasoning traces.
  3. Post-Training: Apply standard methods such as SFT and, where useful, RL.
  4. Inference Loop: Pair the specialist model with an inference loop that generates, evaluates, and improves candidate answers.

Competitive Programming: IOI Specialization

To achieve gold-level performance in competitive programming, NVIDIA curated 22,000 problems and generated synthetic reasoning traces. Two specialist models were developed:

  • Nemotron-3-Nano-CC: A model with 30 billion total parameters (3 billion active) that received both SFT and RL. In IOI 2025 experiments, it improved from 130 points (base) to 291 points after RL, eventually reaching 468 points using the GenCorrect iterative strategy.
  • Nemotron-3-Ultra-CC: A model with 550 billion total parameters (55 billion active) that received SFT. One SFT epoch was sufficient to outperform the fully post-trained Nano model across IOI, ICPC, and LiveCodeBench Pro. This model scored 535.4/600 in IOI 2026.

The findings indicate that adaptation requirements differ by scale; SFT provided the bulk of the gains for the Nano model, while the larger Ultra model required less post-training to reach superior performance.

Olympiad Mathematics: IMO Specialization

Starting from Nemotron 3 Ultra, NVIDIA trained two specialists using a curated SFT corpus of 414,890 quality-filtered examples across 15,818 unique proof problems. The training focused on proof generation, refinement, verification, and meta-verification to teach the model to identify gaps and judge proof completeness.

An RL model was further trained on 9,597 proof problems at the model's capability frontier. The final IMO system combined the general model, the SFT specialist, and the RL specialist in a generate-verify-refine system. This system operated entirely in natural language without external tools, formal provers, or internet access, scoring 30/42 points.

Synergy Between Fine-Tuning and Test-Time Compute

Gold-level results were not achieved through fine-tuning or brute-force sampling alone, but through the co-design of the model, data, and the inference loop. Better specialization provides the inference system with higher-quality candidates, more effective critics, and better refinements.

  • IOI: GenCorrect amplified the gains from fine-tuning through multiple feedback rounds.
  • IMO: Using complementary SFT and RL checkpoints proved more valuable than increasing the sample size from a single checkpoint.

Open Resources

NVIDIA has released several resources on Hugging Face and GitHub to support the community:

  • IMO 2026 Collection: Includes SFT and RL checkpoints, training datasets, and the Nemotron-IMO-Bench (200 olympiad-level problems).
  • NeMo-Skills Repository: Contains the IMO inference pipeline, prompts, and submitted proofs.
  • Competitive Coding: The Nemotron-3-Ultra-CC model is available on Hugging Face, and the IOI paper provides the GenCorrect methodology and training recipe.

Sources