🔬Nemotron 3 Ultra Scores 30/42 for Gold at IMO 2026
TL;DR
NVIDIA says a fine-tuned Nemotron 3 Ultra system scored 30/42 on IMO 2026 problems against a gold threshold of 29, and 535.4/600 on an unofficial IOI 2026 run. The results are self-reported, and the authors credit the inference loop as much as the training.
NVIDIA says a fine-tuned Nemotron 3 Ultra system scored 30/42 on IMO 2026 problems against a gold threshold of 29, and 535.4/600 on an unofficial IOI 2026 run. The results are self-reported, and the authors credit the inference loop as much as the training.
Key Points
IMO: full credit on 4 of 6 problems, natural language only, no formal prover, tools, or internet; graded by official IMO graders
IOI: 535.4/600 versus a 361.12 gold threshold and a 498.27 top human score, unofficial and outside the ranking
Ultra is 550B total and 55B active parameters; Nano is 30B total and 3B active
IMO SFT corpus: 414,890 examples across 15,818 proof problems; released Nemotron-IMO-Bench with 200 problems
Merging SFT and RL checkpoints helped more than sampling more from one checkpoint
Why It Matters
Open-weight models are now in medal territory on olympiad tasks, but the recipe depends on heavy test-time loops. Budget inference cost, not just model size.
Quick Facts
Comments
Be the first to comment
Enjoyed this article?
Get it daily. 7am. Free. Reads in 5 minutes.
Join 3,566 builders reading daily.