Breaking from NVIDIA, and this one has nothing to do with GPUs.
They just released Kumo Tabular, a new family of tabular foundation models that predict new rows in a single forward pass.
Here's the simple version. You have a spreadsheet with thousands of rows (customers, sales, diagnoses) and you need to predict a value for a brand new row. The old way meant training a model from scratch, tuning hyperparameters, and doing manual feature engineering. Slow and painful.
Kumo Tabular flips that. The model takes your existing labeled rows as context, basically saying 'look at these examples,' then instantly predicts the new row. No training. No tuning. No feature engineering.
If that sounds like few-shot prompting in LLMs, you're right. Same core idea, give the model examples in context and let it infer the pattern. Except here the examples are numbers and tables instead of text.
The model is open source, putting NVIDIA alongside TabPFN and TabICL, the projects that pioneered this approach, but now at a bigger scale.
If your job runs on spreadsheets, get ready. AI just learned to read them without a single lesson.
🔗 Original Source & Reference: https://www.marktechpost.com/2026/09/30/nvidia-releases-kumo-tabular/
Published automatically via FeedMind AI Content Pipeline.
Top comments (0)