01 / The short version
What happened
Nemotron 3 was fine-tuned using supervised learning, reinforcement learning, and iterative feedback to reach gold-medal performance at IOI 2026 and IMO 2026. The IOI run was unofficial and conducted under real contest constraints, while IMO proofs were graded by official judges. Two specialized models were built: a 30-billion-parameter Nano model and a 550-billion-parameter Ultra model. Training data included 22,000 programming problems and 414,890 math examples. Performance improved significantly through fine-tuning and test-time refinement, with the final systems combining multiple checkpoints and iterative reasoning strategies to achieve top results.
See the exact references02 / Key takeaways
What you need to know
- 01
Nemotron 3 was successfully fine-tuned for both competitive programming and olympiad math, achieving gold-level results.
- 02
Specialized models outperformed general models, with iterative feedback and multi-checkpoint strategies proving essential.
- 03
Training data, model size, and inference design were co-developed to maximize performance in high-stakes competitions.
Keep in perspective
What to watch for
The IOI result was unofficial and not part of the official ranking, so its competitive standing should be verified against future official benchmarks.
Go to the source
Exact references
These are the original pages used for this brief. Publisher claims are not independent evaluations.
01Primary source · Hugging FaceRead the original announcementhttps://huggingface.co/blog/nvidia/nemotron-ioi-and-imo-2026AI-generated from the linked source. It can miss context; verify consequential details in the original. How the radar works