Neuron-3-Ultra-CC surpasses highest human score in International Informatics Olympiad, marking a post-training leap and test-time inference

AI research has recorded a new milestone in competitive programming after the “Neuron-3-Ultra-CC” model surpassed the highest score achieved by a human contestant in the 2026 International Informatics Olympiad, under a future evaluation that applied the same time limits, internet-access restrictions, and submission rules imposed on human competitors. This result confirms that the development of advanced algorithmic reasoning models is no longer limited to scaling the initial training phase; it now relies on precise engineering of the post-training stage and on leveraging test-time computation through self-evaluation mechanisms and the generation of multiple solutions.

The research team relied on an end-to-end pipeline that combines the collection and curation of 22 000 competitive programming problems, the generation of synthetic reasoning traces, and the application of supervised fine-tuning and reinforcement learning. This pipeline was applied to two architecture variants: the first, “Neuron-3-Nano-CC”, with 30 billion parameters of which 3 billion are active, combined fine-tuning and reinforcement learning; the second, “Neuron-3-Ultra-CC”, with 550 billion parameters and 55 billion active parameters, used fine-tuning alone. Results showed that the smaller model’s scores in the 2025 Olympiad jumped from 130 points to 291 points after post-training alone, before its capabilities doubled with the introduction of advanced reasoning algorithms.

To achieve competitive superiority, the researchers devised a strategy "Gene Correctthat directs test-time computation with feedback, a mechanism that generates diverse program solutions and performs iterative evaluation and modification of outputs before final code adoption. Thanks to this mechanism, the "Nano" model’s score in the 2025 Olympiad rose to 468 points, surpassing the gold-medal threshold of 438.3 points, while the "Ultra" model achieved 502 points. During the live competition of the 2026 Olympiad, the competition-specific "Ultra" system535.4 points out of 600 points, surpassing the gold-medal benchmark of 361.12 points and the top human contestant who scored 498.27 points, becoming the first artificial-intelligence system to outpace the human leader in the competition’s history.

This achievement carries practical implications for CTOs and software-engineering teams in the Gulf, Egypt, and the Levant. The growing reliance ontest-time computationand iterative refinement means that building software systems capable of solving complex problems does not necessarily require running the largest cloud models at continuous cost; engineering teams can employ medium-sized models that have undergone specialized fine-tuning, granting them momentary processing budgets to verify and self-correct code before deployment. This approach enables local companies and banks to lower cumulative inference costs, increase the reliability of critical software systems, and shift programming-assist tools from generating routine text to efficiently solving complex logical puzzles.