Tabular Foundation Models for Data Science Agents
by DeepLearning.AI and Prior Labs
Use TabPFN to predict on spreadsheets without training a model, then hand it to an LLM agent as a tool.
Overview
Tabular Foundation Models for Data Science Agents is an intermediate DeepLearning.AI short course made with Prior Labs, the company behind TabPFN. It is taught by Frank Hutter (founder and co-CEO of Prior Labs), Eliott Kalfon (Principal Data Scientist) and Tuana Çelik (Developer Relations Lead). The course page lists it as last updated on November 11, 2026, with about 1 hour 32 minutes of material across 8 video lessons, 4 code examples and 2 graded assignments, the graded parts sitting behind DeepLearning.AI Pro. The lessons run in order: what tabular foundation models are; classification, regression and forecasting with TabPFN; interpreting TabPFN predictions; real-time inference with the KV cache; building a data science agent; and how tabular foundation models are built. A graded TabPFN AI coding lab and a quiz close it out. The central idea is that a model pretrained on synthetic tables can predict on a dataset it has never seen in a single forward pass, so you skip per-dataset training and tuning. The agent lesson then gives an LLM agent the pretrained model as a tool so it can spot a prediction task and delegate it. TabPFN's GitHub repository (8.4k stars when checked on November 12, 2026) lists TabPFN-3.5 as the default model, supporting up to 1,000,000 rows and 20,000 features, and cites the TabPFNv2 paper in Nature (January 2025). Note the licensing before you ship: the code is Apache 2.0, but the TabPFN-2.5 to 3.5 weights carry non-commercial licenses.
At a Glance
- Topic
- Models
- Level
- Intermediate
- Format
- Course
- Cost
- Freemium
- Duration
- ~1.5 hours of video and labs, self-paced
- Provider
- DeepLearning.AI and Prior Labs
- Hands-on
- Yes — code/exercises
- Certificate
- None
What You’ll Learn
- ✓How tabular foundation models use in-context learning to predict on unseen datasets without training
- ✓Running classification, regression and time-series forecasting on your own tables with TabPFN
- ✓Interpreting TabPFN predictions so you can explain what drove a given output
- ✓Speeding up repeated predictions on the same training table with KV caching
- ✓Building an LLM data science agent that recognizes prediction tasks and calls TabPFN as a tool
- ✓How tabular foundation models are pretrained on synthetic data and what that implies for their limits
Highlights
- •Taught by the team that builds TabPFN, including Prior Labs founder Frank Hutter
- •One of the few courses that connects LLM agents to structured, spreadsheet-style prediction
- •Short enough to finish in an evening: about 92 minutes, with 4 runnable code examples
- •TabPFN is open on GitHub (8.4k stars on 2026-11-12), but the newest weights are non-commercial, which matters for production use
Who It’s For
Best For
- ✓Data scientists who spend too much time training and tuning gradient-boosted models on small tables
- ✓AI engineers building analytics or data science agents that need reliable numeric predictions
- ✓ML practitioners curious about foundation models outside of text and images
Prerequisites
- •Working Python for the hands-on labs
- •Basic familiarity with supervised learning (classification vs. regression) and pandas-style tabular data
FAQ
What is Tabular Foundation Models for Data Science Agents?
A short DeepLearning.AI course, built with Prior Labs, for Python developers and data scientists who want to use tabular foundation models. You learn how TabPFN predicts on a new table through in-context learning, and finish with an LLM-based data science agent that calls the model as a tool for classification, regression and forecasting.
Is Tabular Foundation Models for Data Science Agents free?
Tabular Foundation Models for Data Science Agents offers free content, with paid options for certificates or premium features.
What level is Tabular Foundation Models for Data Science Agents for?
Tabular Foundation Models for Data Science Agents is aimed at a intermediate audience. Recommended background: Working Python for the hands-on labs, Basic familiarity with supervised learning (classification vs. regression) and pandas-style tabular data.
How long does Tabular Foundation Models for Data Science Agents take?
Expect roughly ~1.5 hours of video and labs, self-paced. Most learners work through it at their own pace.
What will I learn from Tabular Foundation Models for Data Science Agents?
You'll learn: How tabular foundation models use in-context learning to predict on unseen datasets without training; Running classification, regression and time-series forecasting on your own tables with TabPFN; Interpreting TabPFN predictions so you can explain what drove a given output; Speeding up repeated predictions on the same training table with KV caching; Building an LLM data science agent that recognizes prediction tasks and calls TabPFN as a tool; How tabular foundation models are pretrained on synthetic data and what that implies for their limits.
Topics
Sources
This page was written from 2 sources, 1 on domains other than deeplearning.ai.