Nvidia's 550B Nemotron-3-Ultra-CC Tops Human Scores at IOI 2026

SANTA CLARA, Calif. — In a notable benchmark milestone for artificial intelligence in algorithmic problem-solving, Nvidia researchers have reported that their massive coding model, Nemotron-3-Ultra-CC, achieved a score of 535.4 out of 600 on the International Olympiad in Informatics (IOI) 2026 problem suite. According to an arXiv research preprint published by the team, this performance comfortably cleared the 2026 gold medal cutoff of 361.12 points and surpassed the highest human contestant score of 498.27 points. If verified, the evaluation marks the first reported instance of an autonomous AI system outscoring the top individual human competitor in the prestigious international high school programming competition.

Background

The International Olympiad in Informatics (IOI) is widely considered the premier global tournament for pre-university competitive programming. Every year, secondary students selected through rigorous national competitions tackle complex computational problems that require advanced algorithmic design, data structures, and rigorous logic.

Unlike traditional natural language processing benchmarks or standardized multiple-choice tests—many of which frontier language models have rapidly saturated—competitive programming remains one of the hardest frontiers in AI evaluation. IOI challenges demand multi-step reasoning, strict computational efficiency under tight time and memory limits, and the generation of exact code tested against hidden test suites.

While AI labs have previously targeted mathematics and coding tournaments, achieving top-tier performance on unseen competition sets without human assistance has proved elusive. Most previous high-profile AI runs either secured silver- or gold-equivalent point bands or required extensive external test-time iteration that did not mirror the strict, real-time submission constraints under which human students compete.

What Happened

In a research preprint titled around the evaluation, an Nvidia research team led by Aleksander Ficek detailed the training and performance of Nemotron-3-Ultra-CC, a competitive programming (CC) variant of the company's 550-billion-parameter architecture. Nemotron-3-Ultra utilizes a sparse Mixture-of-Experts (MoE) design featuring roughly 55 billion active parameters per token, allowing high-capacity reasoning without incurring the computational latency of a dense 550-billion-parameter pass.

Key aspects of the system's performance and architecture include:

• Prospective Evaluation Under Match Constraints: According to the authors, Nemotron-3-Ultra-CC was evaluated prospectively on the IOI 2026 problem suite under conditions matching those faced by student competitors, including strict submission limits, standard grading runs, and no live internet access.
• Point Totals: The model achieved 535.4 points out of a possible 600. The highest-scoring human contestant in the same problem set logged 498.27 points, with the gold medal threshold established at 361.12 points.
• Post-Training Pipeline: The model was developed using an end-to-end curriculum encompassing 22,000 curated competitive programming problems, synthetically generated reasoning chains, and supervised fine-tuning (SFT).
• Test-Time Generation via GenCorrect: Nvidia paired its models with "GenCorrect," an inference strategy designed to iteratively generate, evaluate, and refine algorithmic candidates prior to submission.
• Scaling Across Footprints: Alongside the 550B flagship, Nvidia evaluated a compact counterpart, Nemotron-3-Nano-CC (30 billion total parameters, 3 billion active). Trained with both SFT and reinforcement learning (RL) and paired with GenCorrect, Nano-CC achieved a score of 468 on the IOI 2025 benchmark, demonstrating that aggressive test-time refinement can lift smaller architectures past gold-medal thresholds as well.

While the figures represent an impressive technical leap, the paper notes that the results remain self-reported by Nvidia's research team and await formal review or third-party replication from independent benchmark organizations and the competitive programming community.

Why It Matters

Nvidia’s reported score signals a meaningful shift in the capabilities of reasoning-centric foundation models. Although Nvidia is best known globally as the dominant provider of AI hardware, the company has increasingly built out an internal research and software suite aimed at enterprise agentic workflows and developer toolchains.

A model capable of solving Olympiad-grade problems demonstrates capabilities far beyond basic code autocompletion. Real-world software engineering increasingly relies on autonomous coding agents that can inspect repositories, generate test suites, and troubleshoot edge cases. The multi-step planning and self-correction displayed by GenCorrect and Nemotron-3-Ultra-CC offer a blueprint for agents that do not simply autocomplete syntax, but actively verify code correctness before execution.

Furthermore, the result intensifies competition among frontier frontier AI labs. Organizations such as OpenAI, Google DeepMind, and Anthropic have historically viewed math and coding Olympiads as central benchmarks for artificial general intelligence (AGI) milestones. Demonstrating superior scores on an uncorrupted, newly released IOI problem set establishes a competitive benchmark that rival labs will now face pressure to replicate or match.

What's Next

Despite the milestone, several questions remain before Nemotron-3-Ultra-CC's score can be treated as definitive:

1. Independent Verification: Because the run was performed internally by Nvidia, independent benchmark platforms (such as Vals AI or official IOI committee observers) will need to inspect the submission logs, runtime overheads, and token budgets to verify that the environment fully mirrored official tournament rules.
2. The Human Context: The preprint reports the top human score at 498.27, but full field data—including whether human scores were tightly clustered and how individual problem difficulty distributions played out across the two days of competition—will provide clearer context on the margin of victory.
3. Practical Deployment: While solving isolated algorithmic problems is a critical milestone, adapting these reasoning methods to ambiguous, large-scale software codebases remains a separate challenge. Industry observers will be watching how Nvidia incorporates the Nemotron-3-Ultra post-training recipe into its commercial NIM microservices and enterprise coding tools.

For now, Nvidia's publication presents a formidable claim in computational reasoning, pointing to an era where frontier AI systems do not just compete alongside top human problem solvers, but set the ceiling for them.