NVIDIA Nemotron 3 Wins Gold at IOI 2026 and IMO 2026
NVIDIA's Nemotron 3 family achieves gold-medal results at IOI 2026 with 535.4/600 and IMO 2026 with 30/42, surpassing human top scores and official gold thresholds. Models, weights, and the new Nemotron-IMO-Bench are openly available on Hugging Face and NeMo-Skills.
Executive Overview
NVIDIA announced via the Hugging Face blog that its Nemotron 3 family achieved gold-medal-level results at the International Olympiad in Informatics (IOI 2026) and the International Mathematical Olympiad (IMO 2026). The models, weights, data, and inference pipelines are openly available on Hugging Face and the NeMo-Skills repository. Nemotron-3-Ultra-CC scored 535.4/600 at IOI 2026, surpassing the gold threshold of 361.12 and the highest human score of 498.27. The IMO system scored 30/42, exceeding the official gold threshold of 29. This marks a significant milestone in AI-driven competitive programming and mathematical reasoning.
📊 Official Technical Specifications & Data Sheet
| Technical Aspect | Confirmed Official Data |
|---|---|
| 💰 Pricing & Usage Cost | Models, weights, and data are available free and open on Hugging Face (Nemotron-3-Ultra-CC, Nemotron Labs IMO 2026 collection, Nemotron-IMO-Bench). No paid subscription tiers are announced for these research releases. |
| 🌐 Platforms & Immediate Availability | Hugging Face (models and collections), NeMo-Skills repository (IOI and IMO inference pipeline, prompts, provided proofs, quick start guide). |
| ⚡ Performance & Speed Benchmarks | IOI 2026: 535.4/600 (gold threshold 361.12, highest human score 498.27). IMO 2026: 30/42 (official gold threshold 29, full marks on 4 of 6 problems). IOI 2025: Nano evolved from 130 → 280 (SFT) → 291 (RL) → 468 (GenCorrect), and Ultra-CC reached 502. |
| 🛡️ Security & Prompt Injection Resistance | No security or prompt injection resistance metrics were mentioned in the official announcement; focus is on competitive performance and mathematical/coding reasoning. |
| 🧠 Context Window | Context window size was not disclosed in the official announcement. The system operates entirely in natural language without external tools or internet access during the IMO phase. |
| 🌍 Arabic Language & Regional Support | No explicit Arabic support mentioned; models are multilingual by nature but official focus is on English-language coding and mathematical reasoning. |
Deep-Dive Features & Architecture
NVIDIA revealed a reusable four-step specialization recipe: start from a strong base Nemotron model, curate domain-specific problems with high-quality reasoning traces, apply standard post-training methods (SFT and RL), then pair the specialized model with an inference loop that generates, evaluates, and improves candidate answers. For the IOI track, the team curated 22,000 problems and generated synthetic reasoning traces to train specialists: Nemotron-3-Nano-CC (30B total parameters, 3B active) received SFT and RL, while Nemotron-3-Ultra-CC (550B total parameters, 55B active) received SFT only. Experiments showed that a single SFT epoch was enough for Ultra to outperform fully trained Nano on IOI, ICPC, and LiveCodeBench Pro.
For the IMO track, the team started from Nemotron 3 Ultra and trained two specialists: one with SFT on a dataset of 414,890 examples filtered from 15,818 unique proof problems covering generation, refinement, verification, and meta-verification; the other with RL on 9,597 proof problems selected near the model's capability frontier. The final system combined specialists with the general model in a generate-verify-refine loop, with a final high-compute selection stage, operating entirely in natural language without formal provers, external tools, or internet.
Benchmark & Competitive Performance
On the IOI 2026 benchmark, Nemotron-3-Ultra-CC scored 535.4/600 versus a gold threshold of 361.12 and the highest human score of 498.27 — a human-beating margin of 37.13 points. On IOI 2025, Nano jumped from 130 points before post-training to 280 after SFT and 291 after RL, then to 468 with GenCorrect, surpassing the gold threshold of 438.3, while Ultra-CC reached 502 points. At IMO 2026, the system achieved 30/42 against an official gold threshold of 29, with full marks on 4 of 6 problems. These numbers place Nemotron among specialized models capable of competing with human elites in competitive programming and mathematical proof.
Industry Impact & Enterprise Adoption
The open release of Nemotron-3-Ultra-CC, the Nemotron Labs IMO 2026 collection (including SFT and RL checkpoints, training datasets, and the new Nemotron-IMO-Bench of 200 Olympiad-level problems), and the NeMo-Skills inference pipeline lowers the barrier for enterprises and researchers to build advanced reasoning systems. Organizations can fine-tune these models for domain-specific tasks such as automated code generation, formal verification, and complex mathematical problem-solving. The availability of a standardized benchmark like Nemotron-IMO-Bench enables consistent evaluation and comparison, accelerating innovation in AI-driven STEM applications. This move reinforces NVIDIA's position in the open-source AI ecosystem and provides a foundation for next-generation enterprise AI solutions.
Conclusion
NVIDIA's Nemotron 3 family has demonstrated gold-medal performance at IOI 2026 and IMO 2026, surpassing human top scores and official thresholds. With open access to models, weights, data, and benchmarks on Hugging Face and NeMo-Skills, NVIDIA empowers the global AI community to advance competitive programming and mathematical reasoning. These results signal a new era where specialized AI models can rival human experts in complex, high-stakes domains.
Media Source: Hugging Face | Official Company Statement: Original Source | Fact Verification & Analysis: AI Tools Oasis
Frequently Asked Questions
Nemotron-3-Ultra-CC scored 535.4 out of 600 points at IOI 2026, surpassing the gold medal threshold of 361.12 points and the highest human score of 498.27 points. The result came from a live run under the same time and internet access constraints applied to human contestants.
The Nemotron 3 Ultra system achieved 30 out of 42 points at IMO 2026, exceeding the official gold medal threshold of 29 points, with full marks on 4 out of 6 problems. The proofs were graded by official IMO graders.
Nemotron-3-Nano-CC has 30 billion total parameters and 3 billion active parameters, trained with SFT and RL. Nemotron-3-Ultra-CC has 550 billion total parameters and 55 billion active parameters, trained with SFT only.
Nemotron-3-Ultra-CC is available on Hugging Face. The Nemotron Labs IMO 2026 collection includes SFT and RL checkpoints, both training datasets, and the new Nemotron-IMO-Bench benchmark of 200 Olympiad-level problems. The inference pipeline is available in the NeMo-Skills repository.
Nemotron-IMO-Bench is a new evaluation benchmark comprising 200 Olympiad-level problems, launched within the Nemotron Labs IMO 2026 collection on Hugging Face to enable the community to uniformly evaluate specialized mathematical models.
AI Tools Oasis Team
Bringing you the latest news and analysis in the world of Artificial Intelligence with accuracy and credibility. Follow us for all updates.
