Blog / Analysis
Nvidia's Nemotron outscored every human at IOI 2026, using 760 GPUs
At the International Olympiad in Informatics this year, Nvidia's Nemotron scored 535.4 out of 600 in a live evaluation. The best human contestant scored 498.27 (Nvidia, arXiv 2609.02849). The IOI has run since 1989 as one of the International Science Olympiads, where secondary school students each solve six algorithmic tasks over two days.
The run was not an official IOI entry and the organisers did not supervise it, so Nvidia reports it as "an unofficial, unsupervised benchmark".
Unfair advantage
The rules were the same. Nvidia's system worked under the same two five-hour sessions, the same ban on internet access, and the same limit of 50 submissions per problem as the contestants. What sat behind each submission was not the same: a peak allocation of up to 760 NVIDIA GB300 GPUs, producing up to 200 candidate solutions per round and 1,000 in the final round, with the code compiled and run to choose the ten worth submitting.
A harder year for the humans
Nvidia used IOI 2025 to develop the system and IOI 2026 for the live run, so both years appear in its results. Setting them side by side shows how much the problem set moved.
| Out of 600 | IOI 2025 | IOI 2026 |
|---|---|---|
| Nemotron, one attempt | 304 | not run |
| Nemotron, five GenCorrect rounds | 502.0 | 521.72 (495.0 to 545.8) |
| Top human contestant | 591.23 | 498.27 |
The top human score fell 93 points between the two years and the gold threshold fell 77, so 2026 was the harder set. The 2025 human figures come from the official IOI results. Neither AI score is one of them. OpenAI reported 533.29 on the 2025 problems (The Decoder), which would have placed sixth in a field of 330. Nvidia reported 535.4 in 2026, two points higher, which would have placed first.
What 760 GB300s would cost to rent
The paper gives no cost, and Nvidia ran this on its own hardware. For a sense of scale, GetDeploying put the median GB300 at $7.85 per GPU-hour on 2 September 2026. At that rate the peak allocation of up to 760 GB300s that Nvidia reports for the live run works out at roughly $5,970 an hour.
The teacher was open-weights
Nvidia did not train the competition model on Nvidia data. It fine-tuned Nemotron on reasoning traces generated by GLM-5.2, an MIT-licensed model from Z.ai, after testing it against DeepSeek V4 Flash, also MIT, on the 2025 problems.
| IOI 2025 Score@1 | Score | Mean output tokens |
|---|---|---|
| GLM-5.2 (teacher) | 66.0% | 85,927 |
| DeepSeek V4 Flash (teacher) | 55.3% | 120,456 |
| Nemotron trained on GLM-5.2 | 59.4% | 84,244 |
| Nemotron trained on DeepSeek V4 Flash | 50.7% | 89,626 |
Both teachers score higher than the students they produced. GLM-5.2 answers a single problem better than the 550B model that went on to outscore the best human at IOI 2026, and you can download it. Nvidia also chose on output length: shorter answers fit more candidates into a fixed inference window, so token efficiency is a selection criterion once you are paying for the search.
Questions people ask
How did Nvidia's Nemotron score at IOI 2026?
Nvidia's Nemotron scored 535.4 out of 600 at IOI 2026, against a gold threshold of 361.12 and a top human contestant score of 498.27. It ran on a peak allocation of up to 760 NVIDIA GB300 GPUs. The figures come from Nvidia's paper, arXiv 2609.02849.
Was Nvidia's IOI 2026 result official?
No. The paper states that the system was not an official IOI contestant and that the run was not supervised by IOI, so the score was not included in the official rankings. Nvidia reports it as an unofficial, unsupervised benchmark.
What does the model score without the search loop?
On the IOI 2025 problems Nemotron scored 304 answering once, against 502.0 after five rounds of the GenCorrect generate-and-select loop. Nvidia did not answer the 2026 problems in a single attempt, so 2025 is the only set where both figures exist.
What was Nvidia's IOI model trained on?
Reasoning traces generated by GLM-5.2, an MIT-licensed model from Z.ai. Nvidia measured GLM-5.2 at 66.0% on IOI 2025 against 55.3% for DeepSeek V4 Flash, also MIT, and used GLM-5.2 data for the live run.
Do you need 760 GPUs to run Nemotron 3 Ultra?
No. Nvidia states a serving floor of 8x B200 with about 1.5 TB of aggregate HBM for the published 550B model, or 8x or more H100, H200, GB200 or GB300 across nodes. The 760-GPU allocation was for generating up to 1,000 candidate solutions per problem inside contest time limits.