NVIDIA says a specialized Nemotron system outscored every human at the 2026 coding olympiad -- but its own paper says the run wasn't an official entry

Two weeks after the 37th International Olympiad in Informatics closed in Tashkent, Uzbekistan, NVIDIA researchers say a specialized system built on their own Nemotron architecture did something no AI system has done before: outscored the single best human contestant on the exact same problem set, under matching time and submission limits. The paper, posted to arXiv on September 2, reports Nemotron-3-Ultra-CC scoring 535.4 out of 600 -- more than 37 points clear of China's Xu Qiwen, whose 498.27 topped the actual human field of 375 contestants and stood well clear of the 361.12 needed for gold.

This isn't the first time an AI system has matched or beaten IOI's medal thresholds -- but the shape of NVIDIA's claim is different from the last one. At IOI 2025, an OpenAI reasoning system -- a general-purpose model, not one fine-tuned for the contest -- scored 533.29, gold-medal range, but still placed sixth among that year's 330 human contestants; five humans outscored it. Nemotron-3-Ultra-CC's claim is a different kind: not clearing the bar behind several humans, but topping the single best human score at the event, using a system built and trained specifically for this one contest rather than a general model pressed into service for it.

The system behind the 535.4 isn't a single forward pass. NVIDIA's researchers built GenCorrect, an iterative five-round loop layered on top of the model: generate up to 200 candidate solutions per problem, cluster them down to 10 representative attempts using diversity and "score-blind" heuristics, submit those for judging, then feed the evaluator's per-subtask feedback back into the next round of generation. Parameters tell only part of the story here -- Nemotron-3-Ultra-CC is a 550-billion-parameter model with 55 billion active at a time, the same base architecture NVIDIA released as an open model in June, fine-tuned specifically for this contest on 22,000 curated competitive-programming problems and 477,642 reasoning traces, with every IOI 2025, ICPC 2025, and LiveCodeBench Pro problem explicitly excluded from that training data.

That specialization comes at a cost the paper doesn't hide. Building the pipeline took substantial training compute up front, and running it took substantial compute at test time too -- up to 200 candidate solutions generated per problem, clustered and refined across five rounds, for every one of the six problems in the contest. A human contestant gets one brain and five hours; NVIDIA's system gets a purpose-built training run behind it and a fleet of parallel generations in front of it. The paper's own limitations section makes the same point about resourcing, almost verbatim: the live result is "a system-level comparison under the same time and submission limits, rather than an equal-resource comparison with human contestants."

GenCorrect isn't new this month, either. Run against IOI 2025's own problem set, the same technique took a smaller Nemotron-3-Nano-CC from 291 points to 468 -- already gold-medal range against that year's 438.3 threshold. This year's Ultra-CC result is what NVIDIA's paper calls the first time a system has topped the single highest human score on a full IOI set, rather than simply clearing the medal bar the way both Nano-CC and OpenAI's 2025 system did.

How that run was actually conducted is where the coverage of it starts to disagree.

The paper is blunt about what its own number does and doesn't establish, in language plain enough to quote directly:

Our system was not an official IOI contestant and the run was not supervised by IOI.

Both conditions that were matched to the real competition -- no internet access, the same time and submission limits as the human contestants -- come from the paper itself. What wasn't matched is oversight: nobody from IOI verified NVIDIA's run in real time the way judges verify a contestant's. NVIDIA didn't have to run the evaluation that way. A year earlier, OpenAI said it had "officially entered the 2025 International Olympiad in Informatics (IOI) online competition track and adhered to the same restrictions as the human contestants, including submissions and time limits" -- a mechanism that got its system graded by IOI's own judges, on the record, at the time of the contest. NVIDIA's paper doesn't explain why its team evaluated Ultra-CC independently afterward instead of entering through that track.

None of that makes the underlying capability jump uninteresting -- a system clearing 535 points on a set that handed out only 31 golds this year is a real result, whatever its supervision status. It does mean the honest version of the claim is narrower than the headline most coverage ran with: a heavily specialized pipeline, built and scored by the same lab that wants the story told, beat a human ceiling on a test nobody outside NVIDIA watched it take in real time. NVIDIA has not said whether Nemotron-3-Ultra-CC's weights, or the GenCorrect harness itself, will be released for anyone else to check.