NVIDIA 发布 Kumo Tabular 表格基础模型，单次前向即可预测新行

NVIDIA has released Kumo Tabular , a new family of tabular foundation models (TFMs) for classification and regression. If you have followed TabPFN or TabICL , the setup will look familiar. The model takes labeled rows as context and predicts new rows in one forward pass. There is no training, no hyperparameter tuning, and no feature engineering.

Kumo Tabular comes in Small, Medium, and Large versions, spanning about 28M to 215M parameters. It runs through NVIDIA’s open-source structured-data-models (SDM) library.

Is it deployable? Yes. Weights ship under the OpenMDW-1.1 license, which permits commercial use. The SDM code is Apache-2.0, and it needs Python 3.11+ and PyTorch 2.7+ , with examples targeting a CUDA GPU.

SDM is a GPU-native library for structured-data foundation models and preprocessing. Besides Kumo Tabular, it ships TabICLv2, Google’s TabFM , and KumoRelational for multi-table data. All models share one in-context learning interface built on a TableTensor container. The library also handles preprocessing, ensembling, and many-class prediction.

Kumo Tabular is a Transformer built around the structure of a table. It uses column, row, and in-context attention, as introduced in TabICL and TabPFN. The pipeline has 3 stages:

Because the context never sees the queries, its keys and values are computed once and reused. The head outputs class probabilities, or 999 quantiles for regression. That gives a point prediction plus an uncertainty estimate.

One more detail matters at scale. Softmax attention spreads thin as the number of keys grows. Kumo Tabular scales each query by a temperature that grows with the log of the key count. The coefficient is learned per attention head, so attention stays sharp on larger tables.

Kumo Tabular is pretrained entirely on synthetic tables sampled from Structural Causal Models (SCMs). A random causal graph links hidden variables through linear maps, small neural networks, trees, or Gaussian processes. The generator also injects messy, real-world patterns: missing values, high-cardinality categories, heavy-tailed targets, and conflicting duplicate rows.

Training ran in 3 stages, similar to TabICLv2. Context grew from 1,024 rows to 60,000 rows, with up to 100 columns. Small, Medium, and Large saw about 35M, 71M, and 137M artificial tables. Classification and regression are trained as separate models. NVIDIA says the training recipe and data generators will be released soon.

With default settings, Kumo Tabular ranks first overall on TabArena with an Elo of 1950. NVIDIA team reports it runs 17x faster than LimiX-2 on a single RTX 6000 Pro. All 3 sizes sit on the accuracy and inference-time Pareto front.

Sources: NVIDIA blog , SDM model docs , Prior Labs docs , LimiX GitHub , TabICLv2 paper . Checked September 30, 2026.

The license row is the real differentiator. TabPFN-3, LimiX-2, and TabFM weights carry non-commercial terms. Kumo Tabular and TabICLv2 are the permissive options, and Kumo Tabular leads the benchmarks NVIDIA reports.

Install the library and pass a DataFrame through TableTensor , following the model card :

# pip install structured-data-models from sklearn.datasets import load_breast_cancer import sdm

df = load_breast_cancer(as_frame=True).frame table = sdm.TableTensor.from_pandas( df=df, stypes=sdm.infer_stypes(df, overrides={"target": "categorical"}), device="cuda", ) model = sdm.models.KumoTabular(task="classification", device="cuda") probs = model( x_context=table[:300].drop_columns("target"), y_context=table[:300, "target"], x_query=table[300:].drop_columns("target"), num_estimators=8, )

The size argument accepts "small" , "medium" , or "large" , and defaults to large .