# Nemotron models hit gold level in IOI and IMO contests

> Fine‑tuned Nemotron‑3 variants scored above gold thresholds in the 2026 International Olympiad in Informatics and International Mathematical Olympiad.

Oossa · 2026-10-07 · https://oossa.com/en/nemotron-models-hit-gold-level-in-ioi-and-imo-contests

Hugging Face reports that two specialized versions of its Nemotron‑3 foundation model achieved gold‑medal scores in the 2026 International Olympiad in Informatics (IOI) and the 2026 International Mathematical Olympiad (IMO). The IOI system, called Nemotron‑3‑Ultra‑CC, earned 535.4 out of 600 points, well above the 361.12 gold threshold and higher than the top human score of 498.27. The IMO system, built from Nemotron‑3‑Ultra with both supervised fine‑tuning (SFT) and reinforcement learning (RL), earned 30 out of 42 points, just over the official gold cut‑off of 29.

## What did the researchers do?

Both projects started with a large Nemotron‑3 base model, collected domain‑specific problems, and applied standard post‑training methods. For IOI they curated 22,000 programming challenges and added synthetic reasoning traces, then applied SFT and, for the smaller Nano model, RL. They also used an inference loop called GenCorrect that generates, evaluates, and refines answers. For IMO they assembled a corpus of 414,890 filtered proof examples covering generation, critique, and verification, then fine‑tuned one checkpoint with SFT and another with RL. The final IMO system combined the two checkpoints with the general model, running a generate‑verify‑refine pipeline to produce and improve proofs.

## Where can others find the work?

All models, datasets, and code are released on Hugging Face. The Nemotron‑3‑Ultra‑CC checkpoint is available for download, and the IOI paper describes the GenCorrect methodology. The IMO project includes the SFT and RL checkpoints, the Nemotron‑IMO‑Bench benchmark of 200 olympiad‑level problems, and a NeMo‑Skills repository with the inference pipeline and prompts. The authors hope the community will reuse the “four‑part recipe” – start with a strong base, curate data, apply SFT/RL, and pair with a feedback‑driven inference loop.

## The facts

- Nemotron‑3‑Ultra‑CC scored 535.4/600 on the 2026 IOI, exceeding the gold threshold of 361.12 and the top human score of 498.27.
- The IMO system scored 30/42 points on the 2026 IMO, just above the official gold threshold of 29.
- IOI fine‑tuning used 22,000 programming problems and synthetic reasoning traces; IMO fine‑tuning used 414,890 proof examples across 15,818 unique problems.
- Nemotron‑3‑Nano‑CC has 30 billion total parameters and 3 billion active parameters; Nemotron‑3‑Ultra‑CC has 550 billion total parameters and 55 billion active parameters.

## Why it matters

For developers, the results show that a single large model can be turned into a specialist that rivals top human competitors with a reproducible fine‑tuning recipe. This means hobbyists and research teams can build high‑performing code‑generation or proof‑writing tools without training a new foundation model from scratch.

## Sources & references

1. [One Model Family, Two Gold-Level Results: Fine-Tuning Nemotron for IOI and IMO](https://huggingface.co/blog/nvidia/nemotron-ioi-and-imo-2026) – Hugging Face, 2026-10-07

Last updated: 2026-10-07