Nvidia's Nemotron-3-Ultra-CC scored 535.4 out of 600 at IOI 2026, finishing above every human contestant at this year's International Olympiad in Informatics.

 

The result was published on 2 September in a paper by five Nvidia researchers, who evaluated the system live during the competition rather than replaying the problems afterwards.

 

The olympiad ran in Tashkent, Uzbekistan, from 9 to 16 August. Nvidia says its system worked under the same time, internet-access and submission limits imposed on the students.

 

Related Coverage

 

Inside the IOI 2026 Scoring Run

 

The highest-scoring human competitor at IOI 2026 finished on 498.27 points. Nvidia's system cleared that mark by 37.1 points.

 

It also sat well clear of the gold-medal cutoff, which the organisers set at 361.12 points for this year's problem set.

 

The researchers state the claim directly in the paper's abstract:

 

"Under the same time, internet-access, and submission constraints as human contestants, it scores 535.4 out of 600, exceeding both the gold threshold of 361.12 and the top human score of 498.27. To our knowledge, this is the first AI system to outscore the highest-scoring human contestant on an IOI problem set."

 

The paper, "Post-Training Language Models for Gold-Medal Performance in Coding Competitions," is credited to Aleksander Ficek, Sean Narenthiran, Mehrzad Samadi, Somshubra Majumdar and Boris Ginsburg.

 

How Nvidia Trained the CC Models

 

Nvidia built two competition variants of its open Nemotron 3 family. Nemotron-3-Nano-CC carries 30 billion parameters with 3 billion active, and Nemotron-3-Ultra-CC carries 550 billion with 55 billion active.

 

Both were specialised on a corpus of 22,000 curated programming problems rather than trained from scratch. The pipeline ran in four stages:

 

  1. Large-scale curation of competitive programming problems and test cases
  2. Generation of synthetic reasoning traces distilled from DeepSeek-V4-Flash
  3. Supervised fine-tuning on 1.2 million traces for Nano and 477,000 for Ultra
  4. Reinforcement learning with executable rewards, applied to Nano only

 

Ultra-CC never received the reinforcement learning stage. The authors say compute constraints ruled it out, which means the record score came from supervised fine-tuning plus inference-time search alone.

 

Where GenCorrect Adds Points

 

The larger share of the gain comes from a test-time strategy Nvidia calls GenCorrect. It generates a spread of candidate solutions, runs them, reads the failures and rewrites across five rounds.

 

Its effect is easiest to read on the 2025 problem set, which the team used to tune the approach before the live run.

 

On IOI 2025, Nano-CC scored 130 points before post-training and 291 after it. GenCorrect lifted the same model to 468, past that year's gold threshold of 438.3, while Ultra-CC reached 502.

 

The same models were also measured against ICPC 2025 problems, where Nano-CC solved 51.0% at Pass@1 and Ultra-CC 57.4%.

 

Limits of the Nemotron Comparison

 

The paper is unusually blunt about what the number does not prove. The authors describe the exercise as a system-level comparison, not an equal-resource one.

 

A teenager in Tashkent had one brain and a laptop. Nvidia had a 550-billion-parameter model sampling and re-testing solutions in parallel across five refinement rounds.

 

The authors list several caveats alongside the headline score:

 

  • The approach requires substantial training and test-time compute
  • Compute limits prevented reinforcement learning on Ultra-CC and blocked exhaustive ablations
  • Results may not generalise beyond competitive programming
  • Human contestants and the model were not resourced equally

 

That last point matters for anyone reading the score as a statement about software engineering. Olympiad tasks are self-contained, automatically graded and free of the ambiguity that dominates production code.

 

Nvidia's Release Plan for NeMo-Skills

 

Nvidia says it will publish the competition Ultra-CC checkpoint together with runnable inference and evaluation recipes through its NeMo-Skills repository.

 

The full training corpus stays private. The authors cite third-party redistribution restrictions on the underlying problem sets, which leaves outside groups able to rerun the system but not to rebuild it from the same data.

 

That gap matters because the claim is prospective. The system was scored during IOI 2026 on problems it could not have seen, and independent replication is the normal way such a result gets confirmed.

 

Systems from OpenAI and Google DeepMind reached gold-medal scores at IOI 2025 without finishing above the best human in the hall. Nvidia's contribution is the crossing point, delivered by an open-weight model family rather than a closed frontier system.