Kumo Tabular

NVIDIA's open foundation model for tabular classification and regression, released September 29, 2026. Given a table whose labeled rows serve as context, it predicts the labels of new rows in one forward pass, with no training, tuning or feature engineering. It comes in three sizes from 28M to 215M parameters, each with separate classifier and regressor weights, and was pretrained only on artificial tables. Each table is sampled from a structural causal model, a random causal graph whose nodes apply randomly drawn functions (linear maps, small neural networks, trees or Gaussian processes), and the generator adds real-world mess on purpose: missing values in several patterns, coarsened features (so duplicate rows can disagree on the label), many-level categorical columns and heavy-tailed regression targets. As in TabICLv2, training ran in three stages, from 1,024-row tables to contexts of 400 to 10,240 rows and finally up to 60,000 rows, with up to 100 columns throughout; the small, medium and large models saw about 35, 71 and 137 million tables. The architecture uses the column, row and in-context attention of TabICL and TabPFN: Fourier-feature cell embeddings, induced self-attention down each column, row attention with rotary positions and four [CLS] readout tokens, then a transformer in which query rows attend only to context rows, so the context's keys and values can be cached and reused. A per-head attention temperature that grows with the logarithm of the number of keys keeps attention sharp on tables much larger than those seen in training. The classifier handles up to 10 classes per pass (the library extends this with error-correcting output codes), and the regressor predicts 999 quantiles.

With default settings Kumo Tabular ranks first overall on TabArena with an Elo of 1950 while running 17× faster than LimiX-2 on a single RTX 6000 Pro. It also places first on BeyondArena (Elo 1418), first overall on TALENT (average ranks 6.67, 3.98 and 4.22 for classification accuracy, log-loss and regression RMSE), and first and second on ScoringBench with the large and medium models. Weights are released under OpenMDW-1.1, which allows commercial use, and run through NVIDIA's new Apache-2.0 structured-data-models library, which also packages TabICLv2, Google's TabFM and NVIDIA's relational Kumo model; the training recipe and table generators are promised for later. The model belongs to the NVIDIA Kumo Structured collection and carries the name of Kumo AI, the enterprise predictive-AI startup NVIDIA bought in June 2026 for more than $400 million, according to The Information. Kumo co-founder Jure Leskovec is among the release post's 13 authors. Amazon's Mitra-v2 and Google's TabFM are the other big-lab tabular models trained only on synthetic tables.