NVIDIA’s Nemotron Scores 535.4 on IOI 2026 Problems, Above Every Official Contestant

NVIDIA researchers report that a competition-tuned version of Nemotron scored 535.4 out of 600 on the International Olympiad in Informatics 2026 problem set. That is 174.28 points above the official gold-medal threshold and 37.13 points above the highest-scoring human contestant.

The result was produced during the competition window, before the problems became public, under the same two five-hour sessions, internet restrictions, grading platform and submission limits applied to contestants. NVIDIA says this is the first AI system known to have outscored the leading human on an IOI problem set.

It was not, however, an official entry. NVIDIA’s paper explicitly describes the experiment as an “unofficial, unsupervised benchmark” and says the run was not supervised by IOI. The model does not appear in the official standings and received no medal.

1. What the 535.4 Score Means

IOI 2026 was held in Tashkent, Uzbekistan, from August 9 to August 16. Its 375 contestants represented 92 IOI members. During the competition, participants faced six algorithmic problems—three on each contest day—with a maximum score of 100 points per problem.

The official results place China’s Qiwen Xu first with 498.27 points. Thirty-two contestants reached the 361.12-point gold threshold, while the silver and bronze thresholds were 303.28 and 228.80 respectively.

NVIDIA’s reported 535.4 would therefore rank above every contestant if inserted into the scoreboard. It represents 89.2% of the maximum possible score, compared with 83.05% for the official winner.

That comparison is meaningful because the experiment was prospective rather than a retrospective run against a public benchmark. According to the technical report, NVIDIA operated the system while the competition was in progress and before the six tasks had been released publicly. This sharply reduces the risk that the model had memorized the specific problems from its training data.

The conditions were matched in several other respects. The model had no internet access but could execute code locally. It used the competition grading system and was restricted to 50 submissions per problem, with submissions generally limited to one per minute. The official rules also gave contestants subtask-level scores after each submission, and the model received that same type of feedback.

There is an important verification boundary. NVIDIA’s X post says the score was “as graded by the IOI team,” but the accompanying paper says the model’s run was not supervised by IOI. The official scoreboard independently confirms the human scores and medal thresholds, not the model’s 535.4. The fairest description is therefore an NVIDIA-reported score obtained through the official grading platform, rather than an IOI-certified competition result.

2. The System Was More Than a Single Model Response

The live system was built around Nemotron-3-Ultra-CC, a competitive-programming adaptation of NVIDIA’s Nemotron-3-Ultra-550B-A55B. The underlying mixture-of-experts model has 550 billion total parameters and activates 55 billion for each token.

NVIDIA’s broader post-training project began with a corpus of 22,000 programming problems drawn from 16 competition families and online platforms spanning approximately two decades. The researchers packaged problems into executable environments containing statements, constraints, tests and reference solutions, then removed environments that produced inconsistent results.

For the general Ultra-CC model, the researchers generated 477,642 supervised fine-tuning traces. A smaller 30-billion-parameter Nemotron-3-Nano-CC model received 1.2 million traces plus reinforcement learning, but NVIDIA did not apply the competitive-coding reinforcement-learning stage to Ultra because of its computational cost.

The live IOI 2026 variant used a further competition-specific adaptation. NVIDIA evaluated GLM-5.2 and DeepSeek-V4-Flash as possible teacher models on the 2025 IOI tasks. GLM-5.2 scored 66.0% with a mean generation length of 85,927 tokens, compared with 55.3% and 120,456 tokens for DeepSeek-V4-Flash. A GLM-trained Ultra-CC variant subsequently reached 59.4% on the 2025 problems, versus 50.7% for the DeepSeek-trained version, so NVIDIA selected GLM-5.2 data for the live system.

The decisive component was GenCorrect, an inference procedure that turns the grader’s subtask feedback into an iterative search process. In each of the first four rounds, the system generated as many as 200 candidate programs per problem, discarded invalid outputs, clustered the remaining programs for behavioral diversity and selected ten representatives for submission.

Scores from those submissions showed which subtasks had been solved. The next round then received accumulated subtask results and selected earlier solutions intended to preserve completed subtasks, target unresolved ones and maintain diversity. This design exploits an important feature of IOI scoring: a contestant’s final score for each subtask is the highest score achieved across all submissions, so later programs can pursue different portions of a problem without erasing earlier progress.

For the fifth and final round, NVIDIA expanded the pool from 200 to 1,000 candidate solutions per problem while still submitting only ten. The system generated 50 problem-specific test generators and validators, filtered them to obtain 100 valid test inputs, executed the candidate programs locally and used a model-generated scoring script to rank them. Across five rounds, the procedure could generate up to 1,800 programs per problem while respecting the official limit of 50 graded submissions.

3. NVIDIA Traded Single-Sample Accuracy for Throughput

The live deployment used a peak allocation of 760 NVIDIA GB300 GPUs. That hardware scale makes the result a comparison under matched time and submission rules, not an equal-resource contest between a teenager and one model instance.

To fit the large candidate batches into the competition window, NVIDIA quantized Ultra-CC to NVFP4 and used an FP8 key-value cache with multi-token prediction set to five. On the 2025 IOI development set, this configuration generated 736.8 tokens per second per GPU—3.7 times the 199.1-token rate of the BF16 baseline.

Quantization reduced single-sample accuracy. The selected configuration scored 52.8% on IOI 2025 at Score@1, compared with 59.4% for BF16, a decline of 6.6 percentage points. NVIDIA accepted that trade because GenCorrect depended on generating and testing a much larger solution pool within a fixed five-hour session.

Earlier experiments illustrate why. On the 2025 problem set, the general Ultra-CC model averaged 343.9 points after its first GenCorrect round and 502.0 after five rounds, gaining 158.1 points from iterative sampling and correction. The smaller Nano-CC system rose from 360.6 to 468.2 over the same five rounds.

The result therefore does not show that one unaided model completion surpassed the best contestant. It shows that a model-centered system—combining specialized training, parallel generation, compilation, clustering, synthetic tests, official grader feedback and substantial hardware—could convert a large amount of inference compute into a higher score under the competition’s external constraints.

4. What Is Confirmed and What Remains Unverified

The official IOI records confirm the six-task format, 600-point maximum, gold threshold of 361.12 and Qiwen Xu’s leading score of 498.27. NVIDIA’s technical report supplies the model score, deployment conditions and system details.

The paper reports only one live attempt. NVIDIA later ran the standard GenCorrect pipeline five more times after the contest, obtaining a mean of 521.72 and a range from 495.0 to 545.8. The live score was 13.68 points above that mean but remained within the observed range. Four of those endpoints could have exceeded the top human score at most, while the reported minimum would have fallen 3.27 points short; NVIDIA does not publish each individual score.

The report was submitted to arXiv on September 2, 2026, and is a preprint rather than a peer-reviewed publication. Its authors acknowledge that the approach requires substantial training and inference compute, that Ultra-scale reinforcement learning was outside their budget, and that the findings may not generalize beyond competitive programming.

Reproducibility is also incomplete. NVIDIA says it plans to release the competition checkpoint and runnable inference and evaluation components through NeMo Skills, but the paper says the full training corpus cannot be redistributed because of third-party restrictions. Until the checkpoint and recipes are available, outside researchers cannot independently reproduce the complete live system from the report alone.

For developers, the clearest technical implication is narrower than “AI beats programmers.” The experiment demonstrates that grading feedback can be used as a powerful test-time signal: generate many distinct programs, select a limited submission set, observe which subtasks pass, and allocate later computation to the remaining gaps. It also quantifies how inference throughput can outweigh lower single-response accuracy when a task permits repeated, scored interaction with an evaluator.

Frequently Asked Questions

Did Nemotron win IOI 2026?

No. It was not an official contestant, does not appear in the standings and received no medal. NVIDIA reports that its unofficial live run scored above every official contestant.

What was the highest official human score?

Qiwen Xu of China placed first with 498.27 out of 600. NVIDIA’s reported model score was 535.4.

Did the model have internet access?

No, according to NVIDIA’s report. It could run code locally and access the competition grading platform, matching the relevant contest restrictions.

Was the model limited to one answer per problem?

No. GenCorrect generated hundreds of candidate programs and used up to 50 official submissions per problem across five rounds.

Can researchers reproduce the result now?

Not completely. NVIDIA says it plans to release the checkpoint and runnable recipes, but the full training corpus will not be distributed because of third-party restrictions.

Sources

• Original NVIDIA AI post on X
• Post-Training Language Models for Gold-Medal Performance in Coding Competitions
• Official IOI 2026 Results
• Official IOI 2026 Contest Rules
• International Olympiad in Informatics: IOI 2026