TL;DR
Nvidia says systems adapted from its Nemotron 3 model family reached gold-level scores at the 2026 International Olympiad in Informatics and International Mathematical Olympiad. The IOI result was an unofficial, unsupervised run and was not part of the official ranking; the IMO proofs were graded by official contest graders.
Nvidia says two systems adapted from its Nemotron 3 model family reached gold-medal thresholds at the 2026 International Olympiad in Informatics (IOI) and International Mathematical Olympiad (IMO). The coding system scored 535.4 out of 600 in an unofficial, unsupervised IOI run, while the mathematics system earned 30 out of 42 points on proofs graded by official IMO graders.
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
For IOI 2026, Nvidia reports that Nemotron-3-Ultra-CC, fine-tuned with supervised learning and paired with its GenCorrect generate-evaluate-refine approach, scored 535.4 points. The score was above the reported 361.12 gold threshold and the top human score of 498.27. Nvidia says the run followed the competition’s time, internet-access and submission constraints, but was not included in the official ranking and was unsupervised.
At IMO 2026, Nvidia’s system scored 30/42, above the stated official gold threshold of 29, and received full credit on four of six problems. It combined general Nemotron 3 Ultra with specialist checkpoints trained using supervised fine-tuning (SFT) and reinforcement learning (RL). The system generated candidate proofs, evaluated and critiqued them, then refined selected attempts. Nvidia says it used natural language only, without a formal prover, external tools or internet access.
The training approaches differed by task. For competitive programming, the team curated 22,000 problems and synthetic reasoning traces; for IMO, it reports an SFT corpus of 414,890 quality-filtered examples covering 15,818 proof problems, plus RL training on 9,597 problems. The company says the IMO system drew on multiple checkpoints because their strengths complemented one another.
How Specialization Shaped Both Scores
The two contests test distinct capabilities: IOI asks contestants to devise algorithms and submit code that passes hidden tests under constraints, while IMO requires rigorous written mathematical proofs. Reaching the relevant score threshold in both is evidence, according to Nvidia’s report, that a shared model family can be adapted to different demanding tasks rather than replaced with a separate foundation model for each.
The reported results also point to a role for inference-time feedback alongside training. At IOI, GenCorrect repeatedly generated, evaluated and revised coding attempts. At IMO, a system combined outputs and critiques from general, SFT and RL models. Nvidia’s account presents the outcome as the product of model adaptation, task-specific data and an inference process—not fine-tuning alone. The distinctions in evaluation matter: the IMO proofs had official grading, while the IOI score was not an official competition placement.
Products Worth Considering
Jackery Solar Generator 1000 v2 with Two 100W Solar Panels, 1024Wh LiFePO4 Battery
Apple iPad 11-inch (A16), 128GB, Wi-Fi, Silver
Kindle Paperwhite Signature Edition 32GB E-Reader, 7-Inch Display, Seaglass Green
As an affiliate, we earn on qualifying purchases.
From Earlier IOI Tests to 2026
Nvidia describes the 2026 work as an extension of experiments on IOI 2025. In those tests, its report says, a smaller Nemotron-3-Nano-CC model rose from 130 points before post-training to 280 after SFT and 291 after RL. With GenCorrect, it reached 468 points, above the reported 438.3 gold threshold. Nemotron-3-Ultra-CC scored 502 with the same test-time method.
The company says those experiments informed the IOI 2026 system. For the IMO project, it trained separate proof-focused SFT and RL checkpoints from Nemotron 3 Ultra. Nvidia reports that the SFT model performed best in the first search round, while the RL model had the strongest result among individual checkpoints; the final system used both. These comparisons and training details come from Nvidia’s report, rather than an independent evaluation presented in the source material.
“Success at both points to something broader.”
— Nvidia, in its Hugging Face report
Limits of the Reported Evaluations
The IOI score was a prospective but unofficial run, not an official ranking result. Nvidia says it followed the contest’s stated constraints, but the provided report does not give an independent audit of the run or explain how its unsupervised status affected comparability with human contestants. The reported comparison with the top human score should be read with that qualification.
The source describes official grading for the IMO proofs, but it does not provide an independent assessment of the full system or detail the graders’ evaluation beyond the score. The report also does not establish how well these methods perform outside the two competitions, what their compute costs were, or whether comparable results can be reproduced by other teams. Those questions remain open based on the material provided.
Released Checkpoints and Benchmarks
Nvidia says its Nemotron Labs IMO 2026 collection on Hugging Face includes the SFT and RL checkpoints, their training datasets and Nemotron-IMO-Bench, a benchmark of 200 olympiad-level problems. Making those materials available gives researchers a basis to inspect the methods and test the models beyond the reported contest result. The supplied report does not specify a release date for every asset or describe an independent replication schedule.
Further evaluation will help establish whether the reported gains transfer to new problems and whether the training and inference recipe is reproducible at a comparable cost. For now, the confirmed scope is the two reported 2026 competition outcomes, with the IOI result unofficial and the IMO submissions graded by official contest graders.
Key Questions
What scores did the Nemotron systems report?
Nvidia reports 535.4/600 at IOI 2026 and 30/42 at IMO 2026. It says both scores were above the respective gold thresholds cited in its report.
Was the IOI result an official competition placement?
No. Nvidia describes the IOI run as an unofficial, unsupervised benchmark that followed the competition’s time, internet-access and submission constraints. It was not included in the official IOI ranking.
How was the IMO system evaluated?
Nvidia says the system’s submitted proofs were graded by official IMO graders. It scored 30 points, including full credit on four of the six problems.
What techniques did the teams use?
The report describes supervised fine-tuning and, for some models, reinforcement learning. The systems also used feedback loops to generate, evaluate, critique and revise candidate code or proofs.
Are the models and training materials available?
Nvidia says its Nemotron Labs IMO 2026 Hugging Face collection includes SFT and RL checkpoints, training datasets and a 200-problem olympiad benchmark. The report does not establish whether every IOI training asset is included in that collection.
Source: rss
