• Signal: NVIDIA reports Nemotron-3-Ultra-CC scored 535.4/600 on IOI 2026 against a 361.12 gold threshold and a 498.27 top human score.
• The IOI run was prospective but unofficial and unsupervised, and it is not part of the official IOI ranking.
• The IMO system scored 30/42 against a gold threshold of 29, a one-point margin, with natural-language proofs only and no formal prover, tools or internet.
• Test-time strategy matters as much as training: on IOI 2025, Nemotron-3-Nano-CC went from 291 after RL to 468 with GenCorrect.
• Checkpoints, datasets and a 200-problem benchmark are reported as released, but the blog names no license, so check each page before use.

NVIDIA reports that one model family cleared gold-level thresholds at both the 2026 International Olympiad in Informatics and the International Mathematical Olympiad. Signal, with caveats: the IOI figure is a self-reported, unofficial run, and the IMO margin is a single point. The training recipes, checkpoints and datasets are reported as public, which is what lets readers argue with the claims rather than take them on trust.

What the IOI 2026 run demonstrated

Nemotron-3-Ultra-CC scored 535.4 out of 600 in a prospective evaluation during IOI 2026, per the authors' arXiv paper. The gold threshold was 361.12 and the top human score was 498.27. The run followed contestant constraints on time, internet and submissions, but it was unofficial and unsupervised and is not part of the official ranking. The authors say it is, to their knowledge, the first AI system to outscore the best human contestant on an IOI problem set.

How the coding models were trained

The coding pipeline used 22,000 curated problems with synthetic reasoning traces. Nemotron-3-Nano-CC, a 30B-total, 3B-active model, received SFT and RL; Nemotron-3-Ultra-CC, at 550B total and 55B active, received SFT only. On IOI 2025, Nano rose from 130 before post-training to 280 after SFT and 291 after RL, then to 468 with the GenCorrect test-time strategy, above the 438.3 gold threshold. Ultra-CC reached 502 with the same strategy.

The IMO result is a thin, graded margin

The IMO 2026 system scored 30 out of 42 with full credit on four of six problems, against an official gold threshold of 29. Official IMO graders scored the proofs, which is a stronger check than self-grading. The system worked in natural language only, with no formal prover, external tools or internet access. A one-point margin means a single grader disagreement on partial credit could change the label, so the result is best read as gold-level, not comfortably gold.

Test-time compute carries much of the gain

The reported numbers show that inference-time structure matters as much as weights. Nano's IOI 2025 score moved from 291 to 468 through GenCorrect alone, a larger jump than RL added over SFT. The IMO system likewise combined general, SFT and RL checkpoints in a generate-verify-refine loop. The blog says the SFT checkpoint was strongest in the first search round, while RL gave the best single-checkpoint result. Teams reproducing this need the harness as well as the weights.

What is released, and what is not verified

NVIDIA lists an NVFP4 Ultra-CC model card, an IMO collection with SFT and RL checkpoints, both training datasets, Nemotron-IMO-Bench (200 olympiad-level problems) and NeMo-Skills pipelines. The blog does not name a license, and its text gives 550B parameters while the model page lists 335B, which may reflect quantization but is not explained. I found no independent replication of either result; the figures rest on NVIDIA's own papers and blog.

Verdict: Signal, pending replication

This is Signal, not Noise: the thresholds, scores and constraints are stated precisely, the IMO proofs were graded by official graders, and the artifacts are public enough to test. It is not yet an independently confirmed medal. Watch for third-party reruns of the NeMo-Skills pipelines, a license statement, and whether the IOI result survives scrutiny of the unsupervised-run conditions. Olympiad scores also do not predict performance on production engineering work.