NVIDIA tunes open Nemotron 3 models to gold-level IOI and IMO scores and releases the recipe

In a Hugging Face blog post, NVIDIA says it fine-tuned its Nemotron 3 models into systems that reached gold-medal level at both the 2026 International Olympiad in Informatics (IOI) and the 2026 International Mathematical Olympiad (IMO). At IOI, a specialist called Nemotron-3-Ultra-CC scored 535.4 out of 600, above the 361.12 gold threshold and the top human score of 498.27. At IMO, a system combining several checkpoints scored 30 of 42, one point above the gold line, with full credit on four of six problems. IOI rewards code that passes hidden tests under strict limits; IMO rewards rigorous written proofs, so strong results on both support NVIDIA’s argument that Nemotron is a flexible foundation.

NVIDIA published the checkpoints, training datasets, a new 200-problem benchmark called Nemotron-IMO-Bench, papers and inference code on Hugging Face and in its NeMo-Skills repository. Other Nemotron releases include a diarization model (Nemotron 3 Diarization).

The approach combines supervised fine-tuning, reinforcement learning where useful and a loop that generates, checks and revises answers. For IOI, NVIDIA curated 22,000 programming problems; its smaller Nano model went from 130 points on the 2025 test before post-training to 468 with its GenCorrect feedback loop, crossing that year’s gold line. For IMO, the training set held 414,890 filtered examples across 15,818 proof problems, and the system worked in natural language without a formal prover, external tools or internet. NVIDIA concludes the medals came from designing model, data and search loop together rather than from fine-tuning or brute-force sampling alone.

This contrasts with recent closed-lab math announcements about models outsiders cannot use, including OpenAI’s large release (OpenAI math release). Releasing the pieces lets others test the claim that open base models can become elite specialists, and suggests enterprises can build specialized open models for hard domains without training a new foundation model.

This is a vendor writing about its own models. The IOI run was live and under contestant conditions but unofficial and unsupervised, and it does not appear in official rankings. The IMO proofs were graded by official graders, which is stronger evidence. NVIDIA calls the compute substantial without a public figure, a number that matters for whether university labs could reproduce the work.

Olympiads were built to find and encourage talented students. As they become standard AI benchmarks, there is a risk the contests become marketing stages with students as baselines; using results to build better tutoring and verification tools would be a more constructive outcome. More in our NVIDIA coverage and research coverage.