> ## Documentation Index
> Fetch the complete documentation index at: https://docs.priorlabs.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Models

> Compare TabPFN model versions — capabilities, limits, and availability.

| Model             | Description                                                                                                                                               | Limits                                                                        | Publication                                                         |
| :---------------- | :-------------------------------------------------------------------------------------------------------------------------------------------------------- | :---------------------------------------------------------------------------- | :------------------------------------------------------------------ |
| **TabPFN-3-Plus** | All in TabPFN-3, plus [thinking mode](/capabilities/thinking-mode) and native text feature support. Available through the API and enterprise deployments. | 1M rows × 200 features<br />(trade-off: 100K × 2K, 1K × 20K)<br />160 classes | [Arxiv (2026)](https://arxiv.org/abs/2605.13986)                    |
| **TabPFN-3**      | Latest foundation model. Up to 1M rows in-context with a row/feature trade-off.                                                                           | 1M rows × 200 features<br />(trade-off: 100K × 2K, 1K × 20K)<br />160 classes | [Arxiv (2026)](https://arxiv.org/abs/2605.13986)                    |
| **TabPFN-2.6**    | Good general-purpose choice for datasets up to 100K rows. Previous default model.                                                                         | 100K rows × 2K features<br />10 classes                                       | —                                                                   |
| **TabPFNv2**      | Published in Nature. First TabPFN to handle real-world messy data — mixed types, missing values, categorical features. Best suited for smaller datasets.  | 10K rows × 500 features<br />10 classes                                       | [Nature (2025)](https://www.nature.com/articles/s41586-024-08328-6) |

\* More information about our license can be found [here](/models#tabpfn-model-license).

***

## TabPFN-3-Plus

TabPFN-3-Plus is the API offering built on TabPFN-3. It supports a row/feature trade-off — 1,000,000 rows × 200 features, 100,000 rows × 2,000 features, or 1,000 rows × 20,000 features. Handles numerical, categorical data and missing values automatically. Classification supports up to 160 classes natively (use the [many-class extension](/capabilities/many-class) for higher cardinality). Regression shows up to 20% metric improvements. Predictions run in under 0.2ms on 1M rows.

On top of TabPFN-3, Plus adds two capabilities:

* [Thinking mode](/capabilities/thinking-mode) — test-time compute scaling that spends additional effort at fit time and delivers up to 15% accuracy improvement (measured on TabArena Elo), outperforming AutoGluon 1.5 extreme in less than a tenth of its runtime.
* **Native text features** — pass text columns (product descriptions, customer notes, etc.) directly without encoding or preprocessing. Composes with thinking mode in a single call.

Available through the client and API, as well as enterprise deployments on [AWS SageMaker](/integrations/sagemaker) and [Azure AI Foundry](/integrations/foundry).

***

## TabPFN-3

TabPFN-3 is the default model in the TabPFN OSS package. It has the same foundation as TabPFN-3-Plus — up to 1M rows, 160-class classification, sub-millisecond predictions — but runs locally without the API. Thinking mode and native text features are not available in the OSS package.

The model is licensed under [TABPFN-3 License v1.0](https://huggingface.co/Prior-Labs/tabpfn_3/blob/main/LICENSE). Code under [Prior Labs License](https://raw.githubusercontent.com/PriorLabs/TabPFN/49394b053a6759cfe68e90c21a2d51c31b396768/LICENSE), open source, commercial use with attribution.

***

## TabPFN-2.6

TabPFN-2.6 was the default model in TabPFN OSS package versions v7.0.0–v7.1.x. It shares the same architecture and dataset constraints as TabPFN-2.5 with improved benchmark performance.

* Scales to **100,000 samples and 2,000 features**.
* Handles mixed data types (numerical and categorical) and missing values automatically.

The model is licensed under [TABPFN-2.6 License v1.0](https://huggingface.co/Prior-Labs/tabpfn_2_6/blob/main/LICENSE). Code under [Prior Labs License](https://raw.githubusercontent.com/PriorLabs/TabPFN/49394b053a6759cfe68e90c21a2d51c31b396768/LICENSE), open source, commercial use with attribution.

***

## TabPFNv2

TabPFNv2 is the model published in *Nature* and the first TabPFN to handle real-world messy data — mixed types, missing values, and categorical features automatically.

* Scales to **10,000 samples and 500 features**.
* Best suited for smaller, well-scoped datasets.

The model is licensed under [Prior Labs License](https://raw.githubusercontent.com/PriorLabs/TabPFN/49394b053a6759cfe68e90c21a2d51c31b396768/LICENSE), open source, commercial use with attribution.

<Card title="Nature Publication" icon="book" horizontal href="https://www.nature.com/articles/s41586-024-08328-6">
  Read about TabPFNv2 in our Nature publication from 2025.
</Card>

***

## TabPFN Model License

<Accordion title="What does the non-commercial license mean for enterprise use?">
  Enterprises can download and experiment with the model internally without cost. This includes benchmarking, testing on internal datasets, exploring capabilities, and understanding fit for use cases. Any competitive benchmarking (i.e. for procurement), testing in deployment in production environments, use in workflows that support business decisions, client projects, or revenue generating activities requires a separate commercial license or API agreement.
</Accordion>

<Accordion title="Can enterprises use TabPFN-2.6 or TabPFN-3 for internal testing?">
  Yes. Enterprises may run preliminary internal assessments on their own data for research and exploration. This includes testing on internal datasets, exploring capabilities and understanding fit for use-cases.
</Accordion>

<Accordion title="What is not allowed in enterprise testing?">
  You cannot use the model or its outputs for any activity that influences business decisions, evaluation of vendors, procurement decisions, or commercial workflows. This includes proof of value that affects internal budgeting or vendor selection. In those cases, a commercial license is required.
</Accordion>

<Accordion title="Can we use results from testing to inform product features or deployment strategy?">
  Not without a commercial license. Any evaluation that influences commercial decisions or internal production planning is considered commercial use.
</Accordion>

<Accordion title="Can we evaluate TabPFN-2.6 or TabPFN-3 for integration into our product?">
  You can run small scale internal experiments. If you intend to integrate TabPFN into a product or platform, the API or a commercial license is required.
</Accordion>

<Accordion title="If we want to move from testing to production, what is required?">
  You will need a commercial license or API agreement. Contact [sales@priorlabs.ai](mailto:sales@priorlabs.ai).
</Accordion>

<Accordion title="Do non-commercial rules also apply to derivative models or fine tuned versions?">
  Yes. The same restrictions apply to all derivatives. Fine tuned, modified, or otherwise adapted models remain non-commercial unless you secure a commercial license.
</Accordion>

<Accordion title="What about outputs generated by the model?">
  Outputs are yours, but they can only be used for non-commercial evaluation and research. Using them in production or for any business decisions requires a commercial license.
</Accordion>

<Accordion title="If we train a model from scratch without the weights, is that covered by this license?">
  If you do not use any TabPFN model weights, checkpoints, or other licensed model elements, then the non-commercial license does not apply to that separately developed model. However, if you use TabPFN (or its outputs) to train, guide, fine tune, or distill another model for commercial use, that is not permitted under this license.
</Accordion>

<Accordion title="Can we use TabPFN to train a commercial model indirectly?">
  No. Using TabPFN (or its outputs) to train, fine tune, distill, or otherwise improve a commercial model is not allowed without a commercial agreement.
</Accordion>

<Accordion title="Where should we go if we want to deploy TabPFN in production?">
  Use our API or obtain a commercial license. The API is the simplest path for production trials and product integration.
</Accordion>

<Accordion title="What about previous TabPFN releases?">
  Prior versions remain available under their original licenses. This non-commercial license applies only to TabPFN-2.6 and TabPFN-3.
</Accordion>

<Accordion title="Who do we contact for commercial licensing?">
  Please reach out to our commercial team at [sales@priorlabs.ai](mailto:sales@priorlabs.ai).
</Accordion>