{"version":1,"type":"story","url":"https://digestai.news/story/nvidia-releases-kumo-tabular-model-for-tabular-prediction-with-open-we","json":"https://digestai.news/story/nvidia-releases-kumo-tabular-model-for-tabular-prediction-with-open-we.json","markdown":"https://digestai.news/story/nvidia-releases-kumo-tabular-model-for-tabular-prediction-with-open-we.md","slug":"nvidia-releases-kumo-tabular-model-for-tabular-prediction-with-open-we","headline":"NVIDIA releases Kumo Tabular model for tabular prediction with open weights","summary":"NVIDIA has launched **Kumo Tabular**, an open-source foundation model for tabular data prediction, available on Hugging Face. The model handles classification and regression tasks without training, tuning, or feature engineering, using in-context learning on labeled tables. It comes in three sizes (28M to 215M parameters) and is pretrained exclusively on artificial data generated via procedural sampling to mimic real-world imperfections like missing values and outliers.\n\nKumo Tabular achieves state-of-the-art performance on four benchmarks—**TabArena, BeyondArena, TALENT, and ScoringBench**—while running up to **17x faster** than competing models like LimiX-2 on an RTX 6000 Pro. The model supports up to 10 classes natively and includes built-in preprocessing for numerical/categorical data. NVIDIA provides an open-source library for inference, with weights hosted on Hugging Face under the **OpenMDW-1.1 license**, allowing commercial use. The company emphasizes validation on held-out data due to potential accuracy degradation in edge cases.","keyPoints":["Kumo Tabular predicts labels in a single forward pass with no training, supporting classification and regression on tabular data","Ranks first on four benchmarks (TabArena, BeyondArena, TALENT, ScoringBench) and runs 17x faster than LimiX-2 on NVIDIA GPUs","Pretrained on 35–137 million artificial tables, with weights open-sourced under OpenMDW-1.1 for commercial use"],"whyItMatters":"Kumo Tabular could simplify tabular prediction for industries relying on structured data (finance, healthcare, logistics) by eliminating manual feature engineering and retraining. Its speed and accuracy gains may reduce costs for enterprises using gradient-boosted trees.","category":{"slug":"research","name":"Research","url":"https://digestai.news/category/research"},"entities":{"companies":["NVIDIA","Hugging Face"],"models":["Kumo Tabular","LimiX-2","TabICL","TabPFN"],"people":["David Holzmüller","Vignesh Kothapalli"]},"firstPublishedAt":"2026-09-29T15:30:38Z","updatedAt":"2026-09-29T15:30:38Z","sourceCount":1,"hasPrimarySource":true,"sources":[{"outlet":"Hugging Face","title":"NVIDIA Kumo Tabular Sets a New Accuracy-Efficiency Frontier for Tabular Prediction","url":"https://huggingface.co/blog/nvidia/kumo-tabular","publishedAt":"2026-09-29T15:30:38Z","type":"primary","primary":true,"lead":true}],"sourceNotes":null,"discussions":[],"thread":null,"cite":{"text":"Digest AI, \"NVIDIA releases Kumo Tabular model for tabular prediction with open weights\", 29 September 2026, https://digestai.news/story/nvidia-releases-kumo-tabular-model-for-tabular-prediction-with-open-we","publisher":"Digest AI","title":"NVIDIA releases Kumo Tabular model for tabular prediction with open weights","datePublished":"2026-09-29T15:30:38Z","url":"https://digestai.news/story/nvidia-releases-kumo-tabular-model-for-tabular-prediction-with-open-we"},"generatedBy":"Written by Digest AI's editorial model from the linked sources; the sources are the record.","license":"Headlines, digests and key points are written by Digest AI and may be quoted with a link to the story page. Linked articles belong to their publishers. Terms: https://digestai.news/terms#reuse"}