Concept card

TabPFN / TabICL — Tabular Foundation Models

As of 2026-10-10

Edited by Cédric Mary, sole editor and publication director — 27 years in the SAP ecosystem. How these pages are produced, and where automation is used.

What is TabPFN / TabICL?

TabPFN (Prior Labs, SAP-owned since July 2026) and TabICL (Inria) predict tabular targets by in-context learning — no training, no tuning. TabPFN-3.5 (15 Sep 2026; Plus, Thinking, Fast) is the current release, and TabPFN-3.5 Plus is GA in SAP AI Core, with tabular orchestration support from November 2026.

Tabular foundation models (TFMs) apply the transformer and in-context learning (ICL) to spreadsheets and database tables. Instead of training a new model for every prediction task, one pretrained network receives the labelled rows of the task as input and predicts the unlabelled rows in a forward pass. TabPFN and TabICL are the two best-known open families; SAP's own SAP-RPT (C165, C237) belongs to the same class. Since 2026 this is no longer a research curiosity for SAP consultants: SAP owns Prior Labs, the company behind TabPFN, and TabPFN-3.5 Plus is generally available in SAP AI Core.

Who builds what

TabPFN is developed by Prior Labs (Freiburg, with offices in Berlin and New York). SAP announced the acquisition on 4 May 2026 and completed it on 17 July 2026, committing more than €1 billion over four years to build a frontier AI lab for structured data; Prior Labs continues as an independent entity inside SAP and keeps its open-source strategy (C102).
TabICL is not a Prior Labs model. It comes from Inria's SODA team with ENS Paris, was published at ICML 2025, and has an open successor, TabICLv2, on GitHub (soda-inria/tabicl).

Why it matters

TabPFN is now an SAP asset: SAP completed the Prior Labs acquisition on 17 July 2026 and made TabPFN-3.5 Plus generally available in SAP AI Core on 15 September 2026.
A tabular foundation model gives a usable baseline for a new prediction task in hours instead of a multi-week feature-engineering and training cycle.
Knowing the family (TabPFN vs TabICL vs SAP-RPT vs gradient boosting) lets you run a fair benchmark instead of defaulting to one vendor's model.

Key points

Tabular foundation models predict unlabelled rows from labelled rows passed as input (in-context learning) — no gradient step, no hyperparameter search.
TabPFN is built by Prior Labs (acquisition announced 4 May 2026, completed 17 July 2026, >€1bn over four years); TabICL is an Inria model (ICML 2025), not a Prior Labs model.
TabPFN versions: v1 (2022/ICLR 2023), v2 (Nature 2025), v2.5 (Nov 2025), TabPFN-3 (May 2026), TabPFN-3.5 (15 Sep 2026: base, Plus, Thinking, Fast).
TabPFN-3.5 documents up to 1,000,000 rows (feature-dependent) and up to 20,000 features (6,000 recommended); ranking claims come from vendor-reported leaderboards.
TabPFN-3.5 Plus is GA in SAP AI Core since 15 Sep 2026; tabular orchestration support is planned from November 2026.
SAP-RPT-1.6 (low-latency/high-throughput) and SAP-RPT-1.6-large (highest quality, larger context, more target classes) are the current SAP-RPT generation, shipped with an RPT Playground API rate-limited to 1,000 requests/hour — SAP-RPT-1.5 is the prior generation, not the current one.
Inference cost, not training cost, is the trade-off that decides a TFM's fit: the full labelled context is processed at prediction time, so latency and compute scale with context size, unlike a pre-trained tree ensemble that scores in milliseconds.
TabPFN's licence (modified Apache, TABPFN-3.0 License v1.0) is separate from SAP AI Core's own service terms — check both before committing to a distribution channel.

Terms used on this page

In-context learning (ICL): Inference-time adaptation in which a pretrained model uses labelled examples supplied in its input to predict new cases, without updating weights.
Prior-data Fitted Network (PFN): A transformer pretrained on synthetic datasets drawn from a prior, so that one forward pass approximates Bayesian prediction for a new dataset — the idea behind TabPFN.
Column-then-row attention: TabICL's design: attention across columns, then rows, builds a fixed-size row embedding before in-context learning, which keeps large tables tractable.
TabPFN-3.5 Plus: TabPFN-3.5 variant with proprietary text processing for text-rich tables; the variant made GA in SAP AI Core.
Tabular orchestration: SAP AI Core workflow that manages context retrieval and model routing for tabular foundation models; announced for TabPFN-3.5 Plus from November 2026.

Sources

Full card available to members. What the full card adds: the full decision framework · the SAP vs Snowflake / Databricks / Fabric comparison · the common pitfalls and their fix · the cheat sheet · the architecture schemas · the code blocks · the facts worth quoting.

Open in the app →