NVIDIA has published a technical post describing how its Nemotron model family was adapted into two separate variants, one specialized in competitive programming for the International Olympiad in Informatics (IOI), the other in mathematical reasoning for the International Mathematical Olympiad (IMO), both claimed to reach gold-medal level performance at the 2026 editions.
The effort follows a trend that began in 2025, when several labs — including OpenAI, Google DeepMind, and others — reported gold-medal level results on the IMO using dedicated reasoning models. NVIDIA applies a comparable approach here, but emphasizes reusing a shared base architecture, Nemotron, subsequently fine-tuned separately for each discipline rather than training two unrelated models from scratch.
The distinction between IOI and IMO is technically meaningful: the former demands producing correct, efficient code under strict time and memory constraints, while the latter relies on rigorous mathematical proofs requiring formal chains of reasoning. The fact that a single model family, through targeted fine-tuning, can reach comparable performance across these two very different cognitive tasks is the central point NVIDIA puts forward.
The Hugging Face post does not fully detail the evaluation methodology or the exact conditions used to compare results against official olympiad grading, which warrants some caution before fully endorsing the claimed equivalence to a human gold medal. Nonetheless, the announcement reinforces the growing role of these academic competitions as reference benchmarks for assessing the reasoning capabilities of large language models, beyond conventional test suites.