NVIDIA Nemotron Claims Gold at Both IOI and IMO 2026 Through Specialized Fine-Tuning

NVIDIA's Nemotron 3 family has achieved a rare double: gold-level scores at both the International Olympiad in Informatics (IOI) and the International Mathematical Olympiad (IMO) in 2026. The IOI specialist, Nemotron-3-Ultra-CC, scored 535.4 out of 600 under live contest constraints — well above the 361.12 gold threshold and even the top human score of 498.27. The IMO system, built from Nemotron 3 Ultra checkpoints with a generate-verify-refine loop, earned 30 out of 42 points, surpassing the official 29-point gold cutoff.

A Reusable Specialization Recipe

Both results emerged from the same four-step playbook: start with a strong Nemotron base, curate domain-specific problems with high-quality reasoning traces, apply supervised fine-tuning (SFT) and reinforcement learning (RL) where helpful, and wrap the specialist in an inference loop that generates, evaluates, and refines answers. For competitive programming, the team prepared 22,000 problems and synthetic traces. The 30-billion-parameter Nemotron-3-Nano-CC gained most of its improvement from SFT, with RL adding a consistent boost. The 550-billion-parameter Ultra-CC needed only a single SFT epoch to outperform the fully post-trained Nano across IOI, ICPC, and LiveCodeBench Pro.

Critically, the IOI run was prospective — executed under the same time, internet, and submission limits as human contestants — while IMO proofs were graded by official judges. Neither result required a new foundation model; specialization of the existing Nemotron 3 line was sufficient.