Nemotron IMO Gold (Nemotron-3-Labs-Ultra-Math)

model
NVIDIA

"An Open Recipe for IMO Gold: Training Nemotron for Olympiad Mathematics." NVIDIA's Nemotron Labs released the full stack behind an ensemble that reached a gold-medal level score at the 2026 International Mathematical Olympiad: two post-trained variants of Nemotron 3 Ultra (550B-A55B) specialized to solve hard problems and to find mistakes in proofs (Ultra-Math-SFT and Ultra-Math-RL), the Nemotron-Math-Proofs-v3 SFT and RL datasets, the Nemotron-IMO-Bench evaluation set, and the test-time-scaling recipe in NeMo-Skills. Weights and data ship under OpenMDW 1.1 for commercial and non-commercial use; the report posted to arXiv on September 9 (Ivan Moshkov, Stephen Ge et al.).

Technical Report (arXiv)Technical Report (PDF, NeMo-Skills recipe)HuggingFace CollectionHuggingFace (Ultra-Math-RL)HuggingFace (Ultra-Math-SFT)HuggingFace (Nemotron-IMO-Bench)HuggingFace (Nemotron-Math-Proofs-v3-RL)

Model Details

Architecture MOE
Parameters 550B
Active params 55B
License OpenMDW-1.1
Base model nemotron-3-ultra

Variants

Name Parameters Notes
Nemotron-3-Labs-Ultra-Math-SFT — Supervised stage on Nemotron-Math-Proofs-v3-SFT
Nemotron-3-Labs-Ultra-Math-RL — RL stage on Nemotron-Math-Proofs-v3-RL; the model deployed in the IMO 2026 ensemble

Paper

arXiv HTML PDF HuggingFace
mathreasoningrlopen-weight

Related