One Model Family, Two Gold-Level Results: Fine-Tuning NVIDIA Nemotron for IOI and IMO
Overview
In 2026, NVIDIA's Nemotron model family has demonstrated that a single foundation can be fine-tuned to achieve gold-medal-level performance in two very different domains: the International Olympiad in Informatics (IOI) and the International Mathematical Olympiad (IMO). This dual success highlights how modern large language models, when specialized with targeted fine-tuning, can handle both competitive programming and advanced mathematical reasoning.
The Base Model: NVIDIA-Nemotron-Labs-3-Competitive-Coding-550B-A55B-NVFP4
The work builds on the NVIDIA-Nemotron-Labs-3-Competitive-Coding-550B-A55B-NVFP4 checkpoint, a text-generation model with 335B active parameters. Released on Hugging Face and updated in September 2026, it is designed specifically for competitive coding tasks. Its key architectural features include:
- 550B total parameters / 55B active parameters (A55B): A sparse Mixture-of-Experts (MoE) design that activates only a fraction of weights per token, balancing capability with inference efficiency.
- NVFP4 quantization: A 4-bit floating-point format optimized for NVIDIA hardware, reducing memory footprint and accelerating inference while maintaining accuracy.
- Competitive coding focus: The model is pre-trained and aligned for algorithmic problem solving, making it a strong starting point for IOI-style tasks.
Two Gold-Level Results
IOI: Competitive Programming
Fine-tuning Nemotron on IOI-style problems produced gold-level results. The model generates correct, efficient solutions across a range of algorithmic challenges, including dynamic programming, graph algorithms, and combinatorics. The base model's competitive-coding specialization likely shortened the fine-tuning required to reach this level.
IMO: Mathematical Olympiad
Using the same model family, researchers fine-tuned Nemotron for IMO-level mathematics. This required adapting the model to produce rigorous proofs and multi-step derivations rather than executable code. The result was gold-level performance on olympiad-style problems, demonstrating that the underlying reasoning capabilities transfer across symbolic domains.
Why This Matters in 2026
By 2026, AI systems are increasingly evaluated on their ability to solve open-ended, reasoning-intensive problems rather than just pattern matching. Gold-level results in both IOI and IMO from a single model family signal that:
- Transferable reasoning: Foundation models can be specialized to multiple high-difficulty domains with fine-tuning.
- Efficient specialization: MoE and NVFP4 quantization make it practical to run and adapt very large models.
- Benchmark convergence: Competitive programming and mathematical olympiad tasks are becoming standard stress tests for frontier models.
Technical Takeaways
- The Nemotron family supports both code generation and natural-language proof writing.
- Fine-tuning for IOI emphasizes correctness, complexity, and edge cases.
- Fine-tuning for IMO emphasizes logical rigor, proof structure, and mathematical notation.
- The NVFP4 checkpoint enables deployment on NVIDIA hardware with reduced memory and latency.
Conclusion
The NVIDIA Nemotron Labs 3 Competitive Coding model shows that one model family can achieve gold-level results in two of the most demanding intellectual competitions. As of 2026, this dual capability underscores the growing generality of fine-tuned foundation models and sets a new bar for AI reasoning benchmarks.