Prior Labs

weights

open(TabPFN-2 to 3.5 checkpoints on the Hub, ungated

data

closed(synthetic prior

code

partial(inference and fine-tuning code

license

Apache-2.0(code)+TABPFN-3-License-v1.0(non-commercial weights license

The package code is Apache-2.0, but every current checkpoint - TabPFN-2.5, 2.6, 3 and the default 3.5 - is under a Prior Labs license that allows testing, evaluation and research only, and extends that limit to the model's outputs. Production use needs a commercial license. Only the older TabPFN-2 weights are Apache-2.0 with an attribution clause. TABPFN-3-License-v1.0 does not permit commercial use, so the license tier is commercial_forbidden.

• https://cdn.jsdelivr.net/gh/PriorLabs/TabPFN@main/LICENSE recorded 2026-09-26

  LICENSE is the Apache License, Version 2.0, with the appendix template and no filled-in copyright line; NOTICE names Prior Labs GmbH.

• https://cdn.jsdelivr.net/gh/PriorLabs/TabPFN@main/README.md recorded 2026-09-26

  README: "The TabPFN-2.5, TabPFN-2.6, TabPFN-3 and TabPFN-3.5 model weights are released under non-commercial licenses" and "TabPFN-3.5 is used by default"; the repository is the "Core implementation for fast and local inference".

• https://docs.priorlabs.ai/cookbook/pretrain_nanotabpfn.md recorded 2026-09-26

  The only pretraining walkthrough Prior Labs documents is a teaching model: "This is a teaching model, not the production TabPFN."

• https://huggingface.co/api/models?author=Prior-Labs&sort=downloads&limit=50 recorded 2026-09-26

  Prior-Labs Hub listing: TabPFN-v2-clf, TabPFN-v2-reg, tabpfn_2_5, tabpfn_2_6, tabpfn_3 and tabpfn_3_5, each tagged license:other.

• https://huggingface.co/Prior-Labs/tabpfn_3_5/raw/main/LICENSE recorded 2026-09-26

  TABPFN-3.5 License v1.0 (revised September 9, 2026) uses the same "solely for your Non-Commercial Purposes" grant.

• https://huggingface.co/Prior-Labs/tabpfn_3/raw/main/LICENSE recorded 2026-09-26

  TABPFN-3 License v1.0 grants a license to use the model "solely for your Non-Commercial Purposes", defined as "use for testing, evaluation, or research not tied to commercial gain, production deployment, or revenue generation".

• https://huggingface.co/Prior-Labs/tabpfn_3/raw/main/README.md recorded 2026-09-26

  TabPFN-3 card: "TabPFN-3 is trained purely on synthetic tabular tasks." and "Inference code can be found at https://github.com/PriorLabs/TabPFN".

Measured on monthly PyPI downloads of tabpfn, which is how the weights are fetched and run; the Hub counts on the checkpoint repositories are much smaller.

• https://pypistats.org/api/packages/tabpfn/recent recorded 2026-09-26

  last_month 150,735 downloads of tabpfn

TabPFN does the work of model selection and tuning for you on each new table, and on the TabArena benchmark it outperforms tuned and ensembled gradient-boosting baselines. It stops at one model: it does not search over or ensemble other learners, prepare features, or cover time series and multimodal data the way AutoGluon does.

• https://arxiv.org/abs/2605.13986 recorded 2026-09-26

  TabPFN-3 abstract: "On the standard tabular benchmark TabArena, a forward pass of TabPFN-3 outperforms all other models, including tuned and ensembled baselines, by a significant margin".

• https://arxiv.org/abs/2609.17895 recorded 2026-09-26

  TabPFN-3.5 abstract: "TabPFN-3.5 sets a new state of the art on standard tabular prediction in TabArena".