NVIDIA Fine-Tunes Nemotron for Gold-Level IOI and IMO Results
NVIDIA has published a Hugging Face blog post describing how it fine-tuned its Nemotron models for two top competitions, the IOI and the IMO. The post's title says the one model family produced "Two Gold-Level Results".
- NVIDIA published a Hugging Face blog post on fine-tuning Nemotron for the IOI and IMO, dated 7 October 2026.
- The post's title claims "Two Gold-Level Results" from one model family.
- A linked collection offers checkpoints, training data and a benchmark from "An Open Recipe for IMO Gold: Training Nemotron for Olympiad Mathematics".
- The collection contains 7 items, and a related text generation model is labeled 335B.
- The available text does not include scores or training details.
What NVIDIA published
The post, titled "One Model Family, Two Gold-Level Results: Fine-Tuning Nemotron for IOI and IMO", appeared on the Hugging Face blog under NVIDIA's organization and is dated 7 October 2026. From the title, the work covers two separate contests: the IOI and the IMO, with the IMO referenced as "IMO 2026" in the linked material. The title presents both results as coming from the same Nemotron model family.
The text available for this report contains only the page's headline and surrounding metadata. It does not include the post's body, so details such as scores, the training method or the evaluation setup cannot be reported here.
Open materials for the IMO work
The post links to a Hugging Face collection described as "Checkpoints, training data and benchmark from 'An Open Recipe for IMO Gold: Training Nemotron for Olympiad Mathematics' (IMO 2026)". The collection holds 7 items and was last updated 26 days before the page was captured.
The page also lists one model among those mentioned in the article. It is tagged for text generation and labeled 335B, and its listing shows an update 21 days earlier. The wording "open recipe" in the collection title suggests that NVIDIA is releasing the materials behind the math work, though the available text does not spell out licensing or usage terms.
What it means
The headline claim, that a single model family reached gold-level results in both a programming contest and a mathematics contest, comes from NVIDIA's own post and has not been checked against independent sources here. If the open checkpoints, training data and benchmark are as complete as the collection title implies, researchers and developers could examine and reproduce at least the mathematics side of the work.
The page also points to other recent NVIDIA posts on Hugging Face, including one on Kumo Tabular dated 29 September 2026 and one on NVIDIA Warp and MjWarp for robotics simulation dated 23 September 2026. Readers who want the actual results and methods will need to consult the full post and the linked collection.
Hugging Face: One Model Family, Two Gold-Level Results: Fine-Tuning Nemotron for IOI and IMO
