Nvidia has achieved a landmark result in competitive programming, with its Nemotron-3-Ultra-CC model scoring 535.4 out of 600 at the International Olympiad in Informatics 2026 — decisively beating the top human competitor’s score of 498.27 and clearing the gold medal threshold of 361 points.
The 550-billion parameter model represents a significant leap in AI coding capabilities. Previous results showed Nemotron-3-Ultra-CC scoring 502 on IOI 2025, and this year’s improvement demonstrates continued progress in training methodologies for competitive programming tasks.
Perhaps equally impressive, Nemotron-3-Nano-CC, a compact 30-billion parameter variant, rose from 130 points to 291 points after post-training. When combined with the authors’ GenCorrect test-time scaling strategy, the nano model achieved 468 points — also above the gold cutoff of 438 and a remarkable result for a model orders of magnitude smaller than its frontier sibling.
The training recipe combines a 22,000-problem dataset, synthetic reasoning traces, supervised fine-tuning, and reinforcement learning. This approach demonstrates that frontier-level competitive programming performance can be achieved through carefully curated training data and RL methodologies, not merely through scale.
For the AI industry, the results signal that models are approaching — or have already surpassed — human expertise in structured problem-solving domains like competitive programming. While real-world software engineering involves far more complexity than Olympiad problems, these benchmarks serve as important indicators of model reasoning capabilities.
The achievement adds to Nvidia’s growing portfolio of AI breakthroughs and underscores the company’s strategic investment in AI reasoning capabilities beyond traditional graphics and accelerated computing applications.