Look into the future with TabPFN-Rel

A stakeholder asks: “What is our projected revenue next month?”

You have the customers, orders, and transactions, but answering means pulling those tables together and building a model that predicts historical outcomes. By the time you have an answer, there may already be three follow-up questions, and hence, more work.

We’re releasing an improved version of TabPFN-Rel to help you spend less time building one-off prediction pipelines to answer these questions.

Coding agents help, but reviews create cognitive load

Snowflake Analyst and Databricks Genie already let people ask questions about their data. Claude Code and Codexes of the world can build a whole predictive pipeline out of one question, however, you still need to check if it produces a useful prediction.

For example, when testing churn predictions on last year’s data, an agent might accidentally join test data with current customer records. However, customers who have left this year, are already marked as such. By leaking information the model wouldn’t have when predicting, the test results will look better than they should - but in the next year, on real production data, the performance might be underwhelming.

Agents will also try many things before arriving at the correct feature combination. That is costly both in tokens and time (which is, again, money).

You also still need to check the target definition and model selection. And when a stakeholder asks a follow-up, the reviews start again.

Give your agent a prediction engine

With TabPFN-Rel, your agent uses your org’s knowledge, business context and data schema to define the task: what to predict, for which customers or other entities, and over what period.

TabPFN-Rel follows the relationships between tables and generates features by aggregating related records. It then uses TabPFN-3.5 to make predictions. The agent can reuse the context from your text-to-SQL workflow without writing a custom feature-generation pipeline, which is much more efficient than letting agent do it’s own thing.

A big advantage of using TabPFN-Rel is the built-in guardrails that protect from agents’ making mistakes typical mistakes like temporal leakage, incorrect processing of relational databases and more.

But the cherry on top is that TabPFN-Rel ranks #1 on RelArena-a, a benchmark comparing predictive performance across relational databases. That means trustable, safe predictions that outperform any other baseline.

Better predictions, less waiting

We evaluated the newest version of TabPFN-Rel with TabPFN-3.5-Plus & TabPFN-3.5 on RelArena-a, which you can read more about in our previous blogpost.

Compared to our previous versions, this release with TabPFN-3.5-Plus gains 120 Elo and cements it’s first place. When ran on local GPUs with TabPFN-3.5, it takes the second place, gaining 190 Elo.

One of the key improvements outside of predictive performance is preprocessing - it is up to 100 times faster than the previous release, measured on RelArena-a. We also measured end-to-end speed - TabPFN-Rel is up to 30 times faster than the next best model.

We also added agent skills to get you started fast, and improved interface for defining tasks.

All these improvements for TabPFN-Rel come in a brand new package and repository (become our stargazer!)

Try it yourself

Run the TabPFN-Rel cookbook. It walks through using TabPFN-Rel for predicting which sellers will stop receiving orders, using multi-table e-commerce dataset. Or, even better, point your agent at the dataset from the cookbook and let it figure things out.

To install:

pip install "tabpfn-rel[api]"


TabPFN-Rel is still evolving - so try it on your own data and tell us where you or your agent got stuck.