{"data":{"id":2319,"lang":"en","title":"Nvidia's Nemotron AI Outscores Every Human at Coding Olympiad IOI 2026","summary":"Nvidia's Nemotron-3-Ultra-CC scored 535.4/600 at IOI 2026, topping the best human's 498.27 — the first AI system to beat every human on a full Olympiad-in-Informatics problem set.","category":"ai","category_label":"AI","published_at":"2026-09-05T18:01:26Z","updated_at":"2026-09-05T18:01:26Z","url":"https://techtools.cz/tools/news/article/2319-nvidia-s-nemotron-ai-outscores-every-human-at-coding-olympiad-ioi-2026","api_url":"https://techtools.cz/news-api/v1/articles/2319?lang=en","markdown_url":"https://techtools.cz/news-api/v1/articles/2319.md?lang=en","image_url":"https://api.alphasignal.ai/api/news/preview/nvidia-s-nemotron-beats-the-best-human-at-the-coding-olympics/v2/card.jpg","word_count":152,"reading_time_minutes":1,"sources":[{"url":"https://alphasignal.ai/news/nvidia-s-nemotron-beats-the-best-human-at-the-coding-olympics","domain":"alphasignal.ai"},{"url":"https://www.techtimes.com/articles/326744/20260905/nvidia-ai-outscored-every-human-ioi-2026-how-gencorrect-made-it-possible.htm","domain":"techtimes.com"}],"translation":{"lang":"cs","id":2392,"title":"AI model Nemotron od Nvidie porazil na programátorské olympiádě IOI 2026 všechny lidi","url":"https://techtools.cz/tools/news/clanek/2392-ai-model-nemotron-od-nvidie-porazil-na-programatorske-olympiade-ioi-2026","api_url":"https://techtools.cz/news-api/v1/articles/2392?lang=cs"},"content":"Nvidia's Nemotron-3-Ultra-CC scored 535.4 out of 600 points at the 2026 International Olympiad in Informatics (IOI) in Uzbekistan, beating the top human contestant's 498.27 under identical competition rules and time limits — the first reported case of an AI system outscoring every human on a full IOI problem set.\n\nNvidia trained two systems for the event: Nano-CC, a 30-billion-parameter model (3B active) fine-tuned with both supervised learning and reinforcement learning, and the larger Ultra-CC (550B total, 55B active), trained via supervised fine-tuning on 22,000 curated competitive-programming problems plus 1.2 million reasoning traces distilled from DeepSeek-V4-Flash.\n\nThe breakthrough leaned on GenCorrect, an inference-time \"generate-check-repair\" loop that produces multiple candidate solutions, tests them against available signals, and iterates on failures until it succeeds or runs out of budget. The technique alone lifted Nano-CC's score from 291 to 468 points on last year's problem set — a roughly 60% gain — setting up Ultra-CC's win this year."}}