Editorial illustration for NVIDIA's Kumo Tabular Predicts New Data Rows in One Forward Pass
NVIDIA Kumo Tabular Predicts Data Rows in One Pass
NVIDIA released Kumo Tabular on its structured-data-models library this week, adding a third name to the small but growing list of tabular foundation models that includes TabPFN and TabICL. The pitch is the same as those predecessors: skip training, skip hyperparameter tuning, skip feature engineering. Hand the model labeled rows as context, and it predicts new rows in a single forward pass.
Kumo Tabular ships in three sizes, Small, Medium, and Large, ranging from roughly 28 million to 215 million parameters. Weights are released under the OpenMDW-1.1 license, which allows commercial use, while the SDM library code itself runs under Apache-2.0. Running it requires Python 3.11 or newer and PyTorch 2.7 or newer, with NVIDIA's own examples built around a CUDA GPU.
The release matters less for the headline model than for what sits underneath it. NVIDIA's SDM library is built as a GPU-native home for structured-data foundation models, bundling Kumo Tabular alongside TabICLv2, Google's TabFM, and the multi-table model KumoRelational under one shared interface. How that interface works, and what Kumo Tabular's own architecture looks like under the hood, is where the detail gets specific.
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.
Why this matters
Licensing has quietly become the real differentiator in tabular foundation models, and NVIDIA just picked a side. TabPFN-3, LimiX-2, and TabFM all ship with non-commercial terms, which rules them out for anyone building a product rather than a paper. Kumo Tabular and TabICLv2 are the only permissive options in this comparison, and NVIDIA says its models lead the benchmarks it reports.
That combination, open license plus competitive accuracy, is what makes this worth a second look rather than a shrug. For teams running classification or regression on structured data, the pitch is concrete: no training loop, no hyperparameter search, no feature engineering, just a forward pass over labeled rows. The three sizes (28M to 215M parameters) give founders a real tradeoff between latency and accuracy instead of a one-size model.
We'd still want independent benchmarks outside NVIDIA's own blog before treating this as settled, and the SDM dependency is worth checking against your existing stack. But if the numbers hold up under outside scrutiny, this changes the calculus for teams who've been stuck choosing between TabPFN's performance and TabPFN's license.
Related Reading
Open Source D&B rebuilds 642 million‑business database after AI agents hit limits
Open Source Box CEO Aaron Levie Says AI Will Drive "Biggest Cybersecurity Upgrade
Open Source Study: 6.7% of Deepfake Requests on Hugging Face Targeted Children
Related Topic NVIDIA and Google Cloud let developers scale AI from prototype to production
Market Trends
Related Topic NVIDIA NeMo powers telco reasoning model for autonomous network workflows
Industry Applications
From Archives SBDC Partnership Launches AI Skills Pilot for Small Businesses
From Archives Perplexity's New Embedding Model Outperforms Voyage With 8x Smaller Vectors
From Archives NVIDIA's SCADA Server SDK Lets Partners Build Storage for GPU Clients
LLMs & Generative AI
From Archives NVIDIA's AI Development Constraint: Isolation as a Scaling Unit
Research & Benchmarks
Common Questions Answered
What are the main advantages of NVIDIA's Kumo Tabular compared to traditional tabular machine learning approaches?
Kumo Tabular eliminates the need for training, hyperparameter tuning, and feature engineering by taking labeled rows as context and predicting new rows in a single forward pass. This streamlined approach significantly reduces the time and expertise required to build tabular models compared to conventional methods that require extensive preprocessing and optimization.
How many model sizes does NVIDIA offer for Kumo Tabular and what is their parameter range?
NVIDIA provides three sizes for Kumo Tabular: Small, Medium, and Large, with parameters ranging from approximately 28 million to 215 million. This range of options allows users to choose the appropriate model size based on their computational resources and accuracy requirements.
What licensing advantage does Kumo Tabular have over other tabular foundation models like TabPFN-3 and TabFM?
Kumo Tabular ships with a permissive open license, making it suitable for commercial product development, whereas competitors like TabPFN-3, LimiX-2, and TabFM use non-commercial licensing terms that restrict commercial use. This licensing advantage, combined with NVIDIA's reported competitive benchmark performance, positions Kumo Tabular as a practical choice for companies building commercial applications.
How does Kumo Tabular fit into the growing ecosystem of tabular foundation models?
Kumo Tabular is the third major tabular foundation model released, joining TabPFN and TabICL in a small but expanding category of models designed specifically for structured data. All three models share the same fundamental approach of accepting labeled rows as context and making predictions in a single forward pass, representing a paradigm shift in how tabular data is processed.
Further Reading
- NVIDIA Kumo Tabular Sets a New Accuracy-Efficiency Frontier for Tabular Prediction - Hugging Face
- NVIDIA/structured-data-models: Foundation Models for Structured Data - GitHub
- nvidia/Kumo-Tabular - Hugging Face
- NVIDIA Releases Open Kumo Tabular Model for Tabular Prediction - Unite.AI
- NVIDIA Kumo Tabular Sets a New Accuracy-Efficiency Frontier for Tabular Prediction - NVIDIA Docs
