NVIDIA Nemotron 3 Clears Gold Thresholds at IOI and IMO 2026
NVIDIA's Nemotron 3 systems beat gold thresholds at IOI and IMO 2026, pairing fine-tuning with generate, verify and refine inference in open workflows.
Summary
On October 7, 2026, NVIDIA announced that Nemotron 3 specializations had reached gold level in both 2026 olympiads. Nemotron-3-Ultra-CC scored 535.4/600 at IOI, above the 361.12 gold cutoff and 498.27 top human result, in a live prospective run with contestants’ time, internet and submission limits. The unsupervised test was unofficial and excluded from rankings. At IMO, official graders awarded a generate-verify-refine system using Nemotron 3 Ultra general, SFT and RL checkpoints 30/42, above 29, with full credit on four of six problems.
For programming, NVIDIA curated 22,000 problems and synthetic traces. Nemotron-3-Nano-CC, with 30 billion total and 3 billion active parameters, used SFT and RL; Ultra-CC, with 550 billion total and 55 billion active, used SFT. On IOI 2025, Nano rose from 130 before post-training to 280 after SFT, 291 after RL and 468 with GenCorrect, clearing 438.3; Ultra-CC reached 502. One Ultra SFT epoch beat fully trained Nano across IOI, ICPC and LiveCodeBench Pro, guiding the 2026 system.
For mathematics, SFT used 414,890 filtered examples spanning 15,818 proof problems; RL used 9,597 frontier problems. SFT led the first search round, RL produced the best single-checkpoint result, and both surpassed the general model. Candidates were generated, scored, critiqued, refined and selected at high compute, entirely in natural language without a formal prover, external tools or internet. NVIDIA released the Nemotron Labs IMO 2026 collection with checkpoints, both datasets and 200-problem Nemotron-IMO-Bench. The IMO paper and NeMo-Skills provide methods, prompts, proofs and a quickstart, while Nemotron-3-Ultra-CC, the IOI paper, GenCorrect recipe, evaluation and inference pipelines are available through Hugging Face and NeMo-Skills.
Positives
- Nemotron-3-Ultra-CC scored 535.4/600 at IOI 2026, exceeding both the 361.12 gold threshold and the 498.27 top human score.
- The IMO system earned 30/42 from official graders, beating the 29-point gold threshold and solving four of six problems for full credit.
- GenCorrect lifted Nemotron-3-Nano-CC from 291 after RL to 468 on IOI 2025, above the 438.3 gold cutoff.
- One SFT epoch enabled Ultra-CC to outperform fully post-trained Nano across IOI, ICPC and LiveCodeBench Pro.
- NVIDIA opened checkpoints, datasets, pipelines, submitted proofs and the 200-problem Nemotron-IMO-Bench through Hugging Face and NeMo-Skills.
Risks & concerns
- The IOI 2026 run was unofficial, unsupervised and excluded from the competition’s official rankings.
- The gold-level results required substantial training and iterative inference compute, not fine-tuning alone.
- RL raised Nano’s IOI 2025 score only from 280 after SFT to 291, a smaller gain than SFT or GenCorrect.
- The IMO workflow required a separate high-compute selection stage after proof generation, scoring, criticism and refinement.