NVIDIA reports that Nemotron 3, specialized with SFT, RL, and generate-verify-refine inference, reached gold-medal level at IMO 2026 (30/42) and IOI 2026 (535.4/600). The company released the checkpoints, training datasets, and inference pipelines on Hugging Face and GitHub to make the recipe reproducible.
An arXiv paper presents an end-to-end post-training pipeline for competitive programming, combining problem curation, synthetic reasoning traces, SFT, and RL. It trains Nemotron-3-Nano-CC (30B-A3B) with SFT plus RL and Nemotron-3-Ultra-CC (550B-A55B) with SFT alone, targeting IOI- and ICPC-level performance.
An arXiv paper presents an end-to-end post-training pipeline for competitive programming, combining problem curation, synthetic reasoning traces, SFT, and RL. It trains Nemotron-3-Nano-CC (30B-A3B) with SFT plus RL and Nemotron-3-Ultra-CC (550B-A55B) with SFT alone, targeting IOI- and ICPC-level performance.
NVIDIA released Nemotron-Labs-3-Competitive-Coding-550B-A55B-NVFP4, a competitive-programming specialist fine-tuned from Nemotron-3-Ultra-550B-A55B. Combined with its GenCorrect test-time compute strategy, it scored 535.4/600 on the IOI 2026 problem set under official contest conditions — above the gold threshold and the top human contestant's 498.27 — reportedly the first AI to outscore the best human on an IOI set. Weights are open for commercial and non-commercial use.
NVIDIA released Nemotron-Labs-3-Competitive-Coding-550B-A55B-NVFP4, a competitive-programming specialist fine-tuned from Nemotron-3-Ultra-550B-A55B. Combined with its GenCorrect test-time compute strategy, it scored 535.4/600 on the IOI 2026 problem set under official contest conditions — above the gold threshold and the top human contestant's 498.27 — reportedly the first AI to outscore the best human on an IOI set. Weights are open for commercial and non-commercial use.
A Two Minute Papers video discusses a new NVIDIA paper on human-in-the-loop learning, arguing that simply imitating human actions is insufficient for AI. The video highlights the paper's findings and implications for AI training.
A Two Minute Papers video discusses a new NVIDIA paper on human-in-the-loop learning, arguing that simply imitating human actions is insufficient for AI. The video highlights the paper's findings and implications for AI training.
GPT 5.6 Pro solved all six problems from the 2026 International Mathematical Olympiad on its first attempt, without human help or steering. Former IMO medallists assessed the work and published a report, marking a notable milestone in AI mathematical reasoning.
GPT 5.6 Pro solved all six problems from the 2026 International Mathematical Olympiad on its first attempt, without human help or steering. Former IMO medallists assessed the work and published a report, marking a notable milestone in AI mathematical reasoning.
NVIDIA released Kumo Tabular, an open foundation model for tabular classification and regression that predicts labels in a single forward pass with no training, tuning, or feature engineering. Weights and code are on Hugging Face and GitHub under the OpenMDW-1.1 license, and it ranks first on TabArena, BeyondArena, TALENT, and ScoringBench. It challenges the two-decade dominance of gradient-boosted trees in enterprise tabular ML.
NVIDIA released Kumo Tabular, an open foundation model for tabular classification and regression that predicts labels in a single forward pass with no training, tuning, or feature engineering. Weights and code are on Hugging Face and GitHub under the OpenMDW-1.1 license, and it ranks first on TabArena, BeyondArena, TALENT, and ScoringBench. It challenges the two-decade dominance of gradient-boosted trees in enterprise tabular ML.