Kumo Tabular

What is NVIDIA Kumo Tabular?

NVIDIA Kumo Tabular is a pretrained foundation model for prediction on single tabular datasets. It is designed for data organized into rows and columns, such as customer records, financial applications, transaction histories, demand data, or product measurements. The model accepts labeled context rows, which provide examples of the target outcome, and unlabeled query rows, which are the records to predict.

This approach is called in-context learning. In this setting, the examples are supplied as part of the inference request rather than used to perform a separate training run. Kumo Tabular can therefore be applied to a new classification or regression task without task-specific gradient training and without conventional hyperparameter tuning for each dataset.

The model is provided by NVIDIA and fits within NVIDIA's Kumo family of structured-data foundation models. Kumo-Tabular 1.0 is the tabular model identity documented for the current release. It is separate from language, image, speech, embedding, and graph-prediction systems.

Primary purpose and practical use cases

Kumo Tabular is intended for predictive tasks where the input is a single table containing numerical and categorical features. Classification predicts a category, while regression predicts a numeric value. Examples supported by the model's intended use include churn prediction, credit-risk scoring, fraud detection, conversion prediction, demand forecasting, price estimation, and duration prediction.

A typical workflow might provide historical customer rows containing features such as account age, plan type, usage levels, and payment history, together with a churn label. New customer rows can then be supplied as query rows. The model produces a prediction based on the supplied examples and the relationships it learned during pretraining.

This design is particularly relevant when many small or medium-sized prediction tasks must be explored quickly, or when building a separate conventional model for every table would add too much engineering overhead. It is also useful for structured-data research because the same model family can be evaluated across different tables without retraining a task-specific model each time.

Model variants and architecture

Kumo-Tabular 1.0 is available in three sizes:

All three variants use the same general three-stage transformer architecture, with differences in model width and depth. The architecture includes cell embeddings, a table encoder, and an in-context-learning transformer. Numerical and categorical cells are converted into learned representations before rows and their relationships are summarized for prediction.

Missing values are handled with context-derived imputation and missingness information. In practical terms, this means the model does not treat an absent value as merely an ordinary category without additional context; it uses information from the supplied table and records that the value was missing.

How NVIDIA says the model was trained

NVIDIA reports that Kumo Tabular was pretrained on synthetic tables generated from structural causal models. The synthetic training data included numerical and categorical features, missing values, and classification tasks with between two and ten classes. According to the model documentation, no real-world data was used for pretraining.

These details are provider-reported characteristics rather than an independent evaluation of model quality. They help explain the model's intended generality, but they do not guarantee that it will outperform a task-specific model on every real-world dataset. Performance can depend on the quality and relevance of the context rows, the feature representation, the target task, and the available GPU resources.

Inputs, outputs, and supported modalities

Kumo Tabular works with structured tabular input rather than natural-language prompts. NVIDIA documents support for tabular arrays or Arrow IPC payloads through its Structured Data Models tooling. The request contains a single instance table with numerical and categorical features, labeled context rows, and prediction rows.

For classification, the model returns predictions and class probabilities. For regression, it returns numeric predictions and quantiles, including a native 999-quantile distribution. The NIM endpoint returns prediction results in JSON format. This is structured prediction output, not text generation.

The model does not support image, audio, video, or text input and output as primary modalities. Related tables and graph relationships are not supported by the tabular model. It is therefore not a general-purpose multimodal model, language model, embedding model, or generative assistant.

Context and input limits

NVIDIA does not specify a fixed model-level context window for Kumo Tabular. The practical capacity is determined by factors such as GPU memory, the shape of the input table, and temporary memory allocations during inference.

NVIDIA gives approximately 512 bytes per cell as a rough memory estimate. This is an operational estimate rather than a guaranteed universal limit: tables with more columns, more rows, or larger categorical representations can require more resources. In addition, the Structured Data Models client currently caps the context sent to an inference server at 10,000 rows.

The 10,000-row client cap should not be interpreted as a conventional token context window. Kumo Tabular operates on rows and cells rather than tokens, and the available capacity may also vary by deployment configuration. No maximum output-token limit is applicable because the model produces predictions, probabilities, and quantiles rather than generated text.

Deployment, license, and pricing

Kumo Tabular is available through NVIDIA NIM and as downloadable weights. NVIDIA documents support for Ampere, Hopper, and Blackwell GPU microarchitectures on Linux. The packaged NIM exposes prediction and session endpoints for synchronous inference.

The downloadable weights are released under the OpenMDW 1.1 license. NVIDIA also provides a model page and API documentation for the NIM deployment path. The NIM package is documented as version 1.0.0-rc2, while the model documentation identifies the model version as Kumo-Tabular v1.0.

No token-based model price is published in the supplied information. A free endpoint is identified as available, while self-hosted inference costs depend on the required GPU infrastructure and deployment arrangement. Consequently, there is no verified per-request, per-token, monthly, or recurring subscription price to report. Infrastructure cost may be the more important economic factor for teams running the downloadable model themselves.

Main strengths and limitations

The most distinctive strength of Kumo Tabular is its in-context workflow for structured prediction. A team can supply labeled examples and query rows without building a separate training and tuning pipeline for every new task. The model also handles both classification and regression, includes numerical and categorical data, and exposes useful probability or quantile outputs rather than only a single hard prediction.

The three model sizes provide a deployment choice between resource requirements and model capacity, although the supplied documentation does not provide comparative accuracy, latency, or cost benchmarks for Small, Medium, and Large. Those trade-offs should therefore be measured on the intended hardware and dataset rather than inferred solely from parameter counts.

There are important boundaries. Kumo Tabular is limited to single-table structured prediction and does not natively support relational-table joins or graph relationships. It is not suitable for text generation, image or video generation, speech, embeddings, or web search. It also has no published fixed context length, so large-table workloads must be planned around GPU memory, input shape, and the 10,000-row client limit.

The model's reasoning and coding capabilities are not separately defined because it is not a general-purpose reasoning or programming model. Tool or function calling is likewise not a documented model capability. Its structured outputs should not be confused with an assistant that can execute external tools.

When to choose Kumo Tabular

Kumo Tabular is a strong candidate when the problem is a single-table classification or regression task and rapid experimentation is more important than constructing a bespoke training pipeline. It is especially relevant when labeled examples can be supplied at inference time, when the dataset contains both numerical and categorical columns, or when probability distributions and regression quantiles are useful to downstream decisions.

It may also be a practical choice for structured-data research, small-data experimentation, and teams that want to deploy a pretrained model through NVIDIA NIM or run downloadable weights on supported NVIDIA GPUs.

Another option may be more appropriate when the data spans multiple related tables, when graph structure is central, or when the application requires text, image, audio, video, embeddings, web search, or tool execution. A conventional task-specific machine-learning model may also be preferable when the team can afford dedicated training and needs carefully optimized performance, predictable resource usage, or benchmarked behavior on a particular production dataset.

For larger workloads, users should compare the Small, Medium, and Large variants on their own hardware. The Large variant has substantially more parameters than Small, but the supplied research does not establish that it is always faster, more accurate, or more cost-effective. The appropriate choice depends on the table size, available GPU memory, latency requirements, and the value of improved predictions.

Bottom line

NVIDIA Kumo Tabular is a specialized foundation model for single-table prediction rather than a general-purpose AI assistant. Its central proposition is to make classification and regression predictions from labeled context rows and query rows without task-specific training or hyperparameter tuning. It supports numerical and categorical data, returns probabilities or quantiles, and can be used through NVIDIA NIM or downloadable weights.

Its specialization is also its main constraint. Teams working with one structured table and supported NVIDIA hardware may find the in-context approach useful, while applications requiring language, multimodal generation, relational reasoning, graph data, or a clearly defined token context window should consider a different model type.