On September 2, five Nvidia researchers posted a paper reporting that a system called Nemotron-3-Ultra-CC scored 535.4 out of 600 on the International Olympiad in Informatics 2026 problem set, generated live while the contest was running. The gold-medal threshold was 361.12. The best human in the room scored 498.27. As far as anyone knows, it is the first time a machine has outscored the top-scoring human on a full IOI problem set.
It is also, by the authors' own account, not a fair fight — and unusually, they say so in print.
What the paper reports
“Post-Training Language Models for Gold-Medal Performance in Coding Competitions” (arXiv:2609.02849), by Aleksander Ficek, Sean Narenthiran, Mehrzad Samadi, Somshubra Majumdar and Boris Ginsburg, describes a specialization pipeline built on two base models. Nemotron-3-Nano-CC is a 30B mixture-of-experts with 3B active parameters, trained with supervised fine-tuning and reinforcement learning. Nemotron-3-Ultra-CC is 550B total, 55B active, trained with SFT only — RL at that scale “exceeds our available compute budget.”
The corpus is 22,000 problems from 16 competition families spanning two decades, expanded into 1.2 million synthetic reasoning traces for Nano and 477,642 for Ultra. The inference method, GenCorrect, runs up to five rounds: generate 200 candidate solutions in parallel, cluster them, submit ten representatives, feed the judge's subtask scores back in, repeat. Five rounds of ten maps exactly onto IOI's cap of 50 submissions per problem.
On IOI 2025, used retrospectively, Nano-CC went from 130 points to 291 after post-training and to 468 with GenCorrect, clearing that year's 438.3 gold line. Ultra-CC reached 502. IOI 2026, held across two five-hour sessions with three problems each, was prospective — the problems did not exist publicly when the system solved them, which forecloses the usual contamination objection.
The fine print on “same constraints”
The paper lists the constraints in full: “internet access was prohibited, local code execution was permitted, and each problem allowed up to 50 submissions with one submission allowed per minute.” Time, connectivity, submissions. What that list leaves out is the story.
During the live deployment Nvidia used “a peak allocation of up to 760 NVIDIA GB300 GPUs.” For the competition run the team raised the final-round generation budget from 200 candidates to 1,000, producing roughly 1,800 candidate solutions per problem across five rounds. The paper's own mean generation length for this model is 84,244 tokens. That is roughly 150 million tokens of reasoning per problem, three quarters of a billion across the six-problem set, against a teenager with one laptop.
The team also quantized the model to NVFP4 for the run, accepting a measured 6.6 percentage-point drop in single-shot accuracy to buy a 3.7x throughput increase. That is the design philosophy in one line: volume of attempts beat quality of any single attempt.
The authors are direct about this in their Limitations section: “The live IOI result should therefore be interpreted as a system-level comparison under the same time and submission limits, rather than an equal-resource comparison with human contestants.”
What the “-CC” is hiding
The paper never expands the suffix. It does make clear that what competed in 2026 was not the general Ultra-CC benchmarked earlier, but a “competition-specific system” with three deliberate changes: the quantization, the expanded final round, and a different SFT teacher.
The general models were distilled from DeepSeek-V4-Flash. The live IOI 2026 system was trained on data generated by GLM-5.2 — and in the paper's own Table 1, GLM-5.2 scores 66.0% on IOI 2025 single-shot, higher than any Nemotron variant reported, including Ultra-CC at 59.4%. Nvidia's gold medal was, in substantial part, taught by two Chinese frontier labs. The paper concedes the point obliquely: RL “does not recover the substantially larger gains provided by SFT due to the strength of the teacher models used.”
One more number matters. After the contest, the team re-ran the standard five-round pipeline on the same problems and got a mean of 521.72 with an observed range of 495.0 to 545.8. The bottom of that range sits below the top human's 498.27. The 535.4 is a single draw from a distribution that does not always win.
There is also a supervision question. AlphaSignal reported the run happened in Uzbekistan “with the International Technical Committee supervising.” The paper's footnote says the system “was not an official IOI contestant and the run was not supervised by IOI,” calling the result an unofficial, unsupervised benchmark. The coverage has already flattened the two into one.
What this predicts, and what it does not
The paper evaluates on IOI 2025, ICPC 2025 and LiveCodeBench Pro. There is no SWE-bench result, no repository-level task, no software-engineering evaluation of any kind. The Limitations section says flatly: “Our findings may not generalize beyond competitive programming.”
The structural reason sits inside GenCorrect. The loop needs a judge that is instant, free, honest and granular. IOI supplies exactly that. On ICPC, where feedback is binary rather than subtask-scored, both models “plateau more quickly than on IOI, suggesting that ICPC's binary feedback provides less information for continued refinement.” Degrade the verifier one notch and the method degrades with it. Most software work has no verifier at all, only a flaky test suite and a product manager changing their mind. As one analysis of the result put it, IOI problems are “hard, closed, self-contained, and machine-checkable” — precisely why they yield to this technique, and precisely why it may not travel.
What to watch
Nvidia will release checkpoints and inference components through NeMo Skills but not the training corpus, citing third-party redistribution restrictions, so independent reproduction will be partial at best. Watch for a per-problem breakdown: the paper publishes no problem names, no per-task scores and no subtask table for IOI 2026. Watch for anyone putting a dollar figure on 760 GB300s for two contest days. And watch whether the generate-verify-refine loop survives a domain where the judge is a human being. AlphaSignal's closing line is the right one: the next honest test is what happens when the judge gets messier.
“The live IOI result should therefore be interpreted as a system-level comparison under the same time and submission limits, rather than an equal-resource comparison with human contestants.”— Aleksander Ficek and colleagues, Authors, Nvidia