Every founder skimming last week's headline saw the same three words under SAP's €1B tabular-model acquisition: models stay open. It's true. It's also the most misread sentence in the deal.
Here is the part that matters before you pip install tabpfn and point it at customers.csv: **you can download every TabPFN weight, but you can only legally use one of them in a product.** "Openly available" is a distribution promise. It is not a commercial license. And for a tabular foundation model whose entire pitch is "predict on your business data without an ML team," the gap between those two things is where a solo founder gets hurt.
The one-line version#
TabPFN-2 — the model from the January 2025 Nature paper — is under the Prior Labs License, which is Apache 2.0 plus an attribution line. You can ship it. Every weight released since — 2.5, 2.6, and the current default, 3 — carries a non-commercial license. The commercially usable weight is also the oldest and the smallest.
What "non-commercial" actually forbids#
This is not the soft, gestural non-commercial clause people assume. The tabpfn-2.5-license-v1.1 text on Hugging Face says the model, its derivatives, and its outputs may not be used for any commercial or production purpose, and then it names them:
- revenue-generating products
- client deliverables
- competitive benchmarking for procurement
- internal commercial decision-making
Read that last one twice. Running the newer weights on your churn table to decide which accounts to call is internal commercial decision-making. The proof-of-concept is fine — the license explicitly permits testing, evaluation, and internal benchmarking. Acting on the result commercially is the line you cross. That distinction is exactly the one the "five lines, no ML engineer" tutorials (ours included) don't stop to draw, so draw it yourself.
The demo is licensed. The decision you make from the demo is not.
The version tax#
The instinct is to reach for the newest, biggest weight. Note what that costs you in usable rights:
| Version | Nominal ceiling | License | Commercial? |
|---|---|---|---|
| TabPFN-2 | ~10,000 rows · 500 features | Apache 2.0 + attribution | Yes |
| TabPFN-2.5 | 50,000 · 2,000 | non-commercial | No |
| TabPFN-2.6 | 100,000 · 2,000 | non-commercial | No |
| TabPFN-3 | up to 1,000,000 rows | non-commercial | No |
Moving from the commercial-OK v2 to a weight you can actually run at scale means moving into non-commercial territory. TabPFN-2.5 posts a 100% win rate over a default XGBoost on small data — a genuinely strong result — but you can't put that win in a product without a paid license. The accuracy is free to admire and not free to sell.
Your three honest options#
- Ship TabPFN-2. Commercial use is permitted. You accept the ~10,000-row / 500-feature cap and a generation-old accuracy. For a lot of founder-scale tables — a signup log, a few thousand customers — that ceiling is fine, and this is the cleanest path.
- Pay Prior Labs. The newer weights and the hosted high-speed inference engine are available under a Commercial Enterprise License (
sales@priorlabs.ai). If TabPFN is load-bearing for your product, this is the intended door, and now it has SAP behind it. - Keep the newer weights in the lab. Use 2.5/2.6/3 for evaluation and internal benchmarking only, and don't let their outputs touch a shipped decision.
And the ceiling that ignores your license entirely#
Even with the rights sorted, TabPFN is a small-to-medium tool by construction. Its attention cost scales as O(n² + m²) in rows and features, so memory and latency climb quadratically — that's the real reason for the row caps, not a marketing choice. And on wide data it can simply lose: on high-dimensional problems like RNA-seq with tens of thousands of features, independent work found TabPFN trailing ordinary logistic regression, and trimming to 500 features to make it fit erased its advantage. If your table is short and wide, this is not your model — license or no license.
None of this makes TabPFN a bad bet. It makes it a specific one. The €1B and the "models stay open" line are both real; so is the non-commercial license stamped on every version you'd actually want to run. Decide which weight you're allowed to use before you write the five lines — not after. For the ground-up primer on what a tabular foundation model even is and when it beats gradient-boosted trees, start with our TabPFN vs XGBoost breakdown, then come back here for the fine print.
