NVIDIA Releases Nemotron-Labs-3 550B Model Specialized for Competitive Coding
NVIDIA has released Nemotron-Labs-3-Competitive-Coding-550B, a massive open-weights model fine-tuned on 477,642 synthetic reasoning traces distilled from GLM-5.2. Paired with a novel test-time compute strategy named GenCorrect, the model scored 535.4 out of 600 on the IOI 2026 problem set, outscoring the highest-scoring human contestant. This achievement marks the first reported instance of an AI system outperforming the top human contestant on an International Olympiad in Informatics (IOI) problem set under contest conditions. It demonstrates how synthetic data distillation and iterative test-time refinement can enable open-weights LLMs to master highly complex algorithmic logic. The 550-billion parameter model is quantized in NVIDIA's native 4-bit floating-point format (NVFP4) for optimized hardware inference and is available for both commercial and non-commercial use. Fine-tuned for one epoch over 22,000 curated problems across 16 contest families, the model used GLM-5.2 as its teacher model because it yielded higher accuracy and roughly 30% shorter outputs than alternative variants.
## BACKGROUND
Competitive programming competitions, such as the International Olympiad in Informatics (IOI), test advanced algorithm design and optimization under strict speed and computational constraints. Test-time compute strategies, like GenCorrect, spend extra compute resources during evaluation to generate candidate code, run automated tests, and iteratively refine solutions based on feedback. NVFP4 is a specialized 4-bit floating-point quantization format designed for NVIDIA Tensor Cores to significantly reduce model memory footprint while preserving precision.