NVIDIA's Tabular AI Needs Zero Training

🔥 What happened NVIDIA dropped Kumo Tabular: tabular foundation models that predict new rows in a single forward pass — no training, no tuning, no feature engineering. Three sizes from 28M to 215M parameters, pretrained purely on synthetic tables. 💡 Why it matters The weights ship under OpenMDW-1.1, which allows commercial use. Rivals TabPFN-3, LimiX-2, and TabFM all carry non-commercial terms. NVIDIA reports first place on TabArena with Elo 1950 and 17x faster inference than LimiX-2 on a single RTX 6000 Pro. ⚡ Our take NVIDIA isn't winning tabular AI with a better architecture — it's winning with the license. If you run TabPFN in production today, you have a compliance problem tomorrow, and NVIDIA has the fix.