Nemotron fine-tuned to gold-medal level on IOI 2026 and IMO 2026 benchmarks

Hugging Face and NVIDIA report that Nemotron models fine-tuned for competitive programming and olympiad mathematics achieved gold-medal level results in IOI 2026 and IMO 2026 benchmarks. The IOI 2026 result was produced by Nemotron-3-Ultra-CC with supervised fine-tuning (SFT) and the GenCorrect inference strategy, scoring 535.4 out of 600, above the 361.12 gold threshold and surpassing the top human score of 498.27. The IMO 2026 result was produced by a generate-verify-refine system using Nemotron 3 Ultra SFT and RL checkpoints, scoring 30 out of 42 points and exceeding the official gold threshold of 29.

The teams used a reusable specialization recipe consisting of four parts: starting with a strong Nemotron base model, curating domain-specific problems and reasoning traces, applying standard post-training methods such as SFT and RL, and pairing the specialist model with an inference loop that generates, evaluates, and improves candidate answers. For competitive programming, two specialists were trained: Nemotron-3-Nano-CC (30B total parameters, 3B active) and Nemotron-3-Ultra-CC (550B total parameters, 55B active). Nano improved from 130 points before post-training to 280 after SFT and 291 after RL, then reached 468 with GenCorrect, crossing the gold threshold of 438.3. Ultra-CC reached 502 points with the same test-time strategy and 535.4 in the live IOI 2026 run under human-like constraints.

For IMO mathematics, the SFT corpus contained 414,890 quality-filtered examples across 15,818 unique proof problems, covering proof generation, refinement, verification, and meta-verification. An RL model was trained on 9,597 proof problems near the capability frontier. The final system combined both SFT and RL checkpoints with a generate-verify-refine pipeline, scoring 30 out of 42 points and earning full credit on four of six problems.

The organizations released open assets to support reproducibility and broader use. The Nemotron Labs IMO 2026 collection includes SFT and RL checkpoints, training datasets, and Nemotron-IMO-Bench, a benchmark of 200 olympiad-level problems. The IMO paper describes the training approach and system, while the NeMo-Skills repository provides the inference pipeline, prompts, submitted proofs, and a quickstart. For competitive programming, Nemotron-3-Ultra-CC is available on Hugging Face, and the IOI paper provides the training recipe and GenCorrect methodology.