The state of Tabular Foundation Models (2026)

Pre-trained models, startups, and recommendation.

This is post #5 of the Tabular Foundation Model (TFM) series (see Part 1, Part 2, Part 3, and Part 4).

While working on this overview of tabular foundation models, I took a break and scrolled LinkedIn (guilty as charged 😅), where I saw that TabICL v2 was just published. This anecdote shows how quickly the TFM space is evolving.

The prior-data fitted networks opened up a new paradigm for supervised machine learning. And this led to rapid innovation with researchers proposing new models. These last few years have actually been quite convenient for the tabular machine learner: It was clear that gradient-boosted tree frameworks like xgboost or Catboost are solidly on the throne.

Enter the — less comfortable — world of tabular foundation models, where SOTA changes every couple of months or weeks. While it may take a while for a clear winner to bubble up, this post offers a snapshot:

• This post gives an overview of tabular foundation models.

• I discuss how TFMs may change tabular ML.

• At the end, I recommend a TFM.

Let’s go.

Quick Timeline

Tabular foundation models are still relatively new.

It all started with the prior-data fitted networks, short PFNs, which I covered in this post. PFNs are a framework that allows pre-training on tabular data without requiring the use of column names or similar information, instead focusing on learning to learn, using attention mechanisms.

TabPFN was published in 2022 as the first PFN-based tabular foundation model. Clunky, only to be used with small data. But a proof of concept. A few other approaches emerged in 2023 and 2024, but it was in 2025 when the field picked up the pace. At the beginning of 2025, TabPFN v2 was published, and IMHO, that’s when TFMs crossed from proof-of-concept to a serious contender in tabular ML. Since then, many TFMs have been published, all with different architectures and pre-training.

Today, TFMs may beat gradient-boosted trees. For example, TabArena compares various supervised machine learning algorithms on various regression and classification tasks. On this leaderboard, RealTabPFN v2.5 is the best TFM. Only AutoGluon extreme (4h) is better, but this is an expensive ensemble of many approaches on TabArena.

Note that TabArena comes with a caveat: Researchers from the same lab that published TabPFN are involved. I have no reason to think anything shady is going on, but it’s just generally better when such an evaluation is done by an independent party. Another independent and relevant benchmark is the TALENT benchmark. The following figure shows the performance of TabICL v2 (and others) on TabArena.

Train time versus improvability for various ML algorithms on the TabNET dataset. Source: TabICL repo.

Which TFMs are available?

So, let’s get to the gist of it: Which tabular foundation models are available right now? Probably more than in the list, but here are the ones I found, ordered chronologically:

• PFN (2021 | paper, repo) — The basis for most foundation models; performs Bayesian posterior predictive inference via transformers.

• TabPFN v1 (2022 | paper, repo) — First ICL transformer for small tabular classification.

• MotherNet (2023 | paper, repo) — Hyper-network transformer pre-trained to set weights of a child neural network directly.

• TabForestPFN (2024 | paper, repo) — Forest-based dataset generator enabling complex decision boundaries; extends TabPFN.

• TabDPT (2024 | paper, repo) — Combines ICL with self-supervised learning on 123 real datasets; shows power-law scaling.

• TabPFN v2 (2025 | paper, repo) — Handles ~10k samples, 500 features, 10 classes using row + column self-attention.

• TabICL (2025 | paper, repo) — Column-then-row attention; scales to ~500k samples; ~10× faster than TabPFN v2 on large dataset with many columns.

• TabFlex (2025 | paper, repo) — Uses linear attention to scale tabular learning to millions of samples.

• ContextTab (2025 | paper, repo) — Semantics-aware ICL trained on real-world tabular data.

• Real-TabPFN / TabPFN v2.5 (2025 | paper, repo) — Continued pre-training on curated real-world datasets to improve performance.

• LimiX (2025 | paper, repo) — Models joint distribution over variables and missingness; introduces LDM scaling law.

• Mitra (2025 | paper, repo) — Uses curated mixtures of synthetic priors and fine-tuning; surpasses TabPFN v2 and TabICL.

• TabPFN-Wide (2025 | paper, repo) — Continued pre-training for extreme feature counts (>50k); targets biomedical settings.

• Orion-MSP (2025 | paper, repo) — Multi-scale sparse attention with Perceiver-style memory for tabular ICL.

• Orion-Bix (2025 | paper, repo) — Bi-axial attention designed for few-shot tabular learning.

• TabICL v2 (2026 | paper, repo) — Improves speed and predictive performance over TabICL v1.

In a way, this list reminds me of the situation of large language models, where we have this explosion of pre-trained models. Each month, we get a new SOTA on this or that benchmark.

Something else is also happening due to TFMs with their pre-training nature: There are many dedicated TFM companies.

The rise of the TFM start-up

TFMs have kicked off a development in tabular machine learning that I haven’t seen before: Startups dedicated specifically to pre-training foundation models. Clearly, we just didn’t have foundation models before, but we didn’t really see startups focused on a single supervised ML algorithm, like an xgboost startup. And this is because of the paradigm-shift that PFNs enabled, from “traditional” supervised ML to the foundation model era. And it kind of makes sense, because all of a sudden, you can pre-train a model for specific industries and scenarios.

There are now quite a few startups:

Besides PriorLabs (TabPFN) in Germany, France has NeuralkAI (NICL) and The Forecasting Company. Looking overseas, there is Layer6 AI (TabDPT) in Canada, as well as Kumo, Fundamental (NEXUS), and Lexsi Labs (Orion-MSP, Orion-Bix) in the USA. Remark: Fundamental seems to have offices in Spain, and Lexsi Labs seems to be hiring mostly in India and might have offices there. For China, I found StableAI (LimiX). Also, some of the big players are training their foundation models: Microsoft (Mothernet), Amazon AWS (Mitra), and SAP (ContextTab).

Many of the foundation models that I listed in the overview can actually be associated with a startup or big tech. This is not entirely new. For example, CatBoost is developed by Yandex.

But that brings me to another big caveat: licensing. The weights of RealTabPFN from Prior Labs, the successor of TabPFN v2, come with a non-commercial license: You are free to try the model out or use it in research, but you have to pay for commercial use.

I think it’s exciting that you can build a business model around the tabular foundation models. At the same time, I fear that it might move tabular ML into a more closed-source direction. Which would be sad, since tabular machine learning is a very open community, embracing open source, open publications, and so on.

Which foundation model to pick?

So many TFM options … which one should I pick?

My personal pick: TabICL v2.

TabICL v2 is fast and highly performant. Also, it’s fully open source. I was very happy about TabICL v2 coming out. Especially seeing that they now support regression, my hopes are high that TabICL emerges (and remains) as a fully open-source SOTA. Take my recommendation with a grain of salt: I’ve only tried out a few models, and also only on toy data. In the end, just try out the models yourself 😉.