TabPFN-3.5-Thinking sets the new bar

Diana Kriuchkova
-
Nick Erickson
-

When you’re forecasting electricity prices for the next bid or deciding how much stock to buy for the holiday season, better predictions can have significant business value. TabPFN-3.5-Thinking uses more compute during training and is made for tasks where greater accuracy is worth the investment.

We are upgrading Thinking mode to outperform every other foundation or traditional model available on the market, now on datasets with up to 1,000,000 rows. New Thinking also entirely defines the Pareto frontier: for every other method there is a Thinking variant that is both more accurate and faster to train.

__wf_reserved_inherit

The Thinking upgrade is live on our API. Try it on your own data in the playground.

Try Thinking in the playground ↗

More accurate predictions, at a larger scale

TabPFN-3.5 already supports up to 1M rows. Today, we bring Thinking to the same scale, up from its previous limit of 200k rows. For energy trading, that means using more observations of price and demand on the market. For retail stock planning, it means learning from more patterns across products, stores and promotions.

And it shows on the BeyondArena-large, covering datasets with 100k to 1M training rows. TabPFN-3.5-Thinking (max) has a winrate of 95.2% against RealMLP, the previous #1 model. It also beats Kumo-Tabular with a winrate of 90.8%, and even outperforms our own TabPFN-3.5.

__wf_reserved_inherit

We also optimized the latency. TabPFN-3.5-Thinking mode now entirely defines the Pareto frontier: for every other method there is a Thinking variant that is both more accurate and faster to train. On large-data tasks, Thinking mode takes 32% less prediction time than the previous version. If you are predicting on large volumes, that means less time waiting and more time to act.

Account for patterns across customers, products and time

The outcome of churn, revenue or a price can depend on the customer, product or device. People also often spend more during a promotion or bonus period, customers might churn after a period of inactivity, and Christmas might increase the demand for a particular product group. This is grouped and temporal data.

__wf_reserved_inherit

Now, TabPFN-3.5-Thinking is the best model for predictions on every type of data - from grouped, to temporal, high-cardinality and high-dimensionality. We measured the performance on BeyondArena, and it consistently took first place on every data subslice, beating the closest rivals - both classic machine learning models like RealMLP and other tabular foundation models.

All in all, Thinking gives you a better basis for planning around the customers, products and periods that matter to your business.

Try the new TabPFN-3.5-Thinking

The new Thinking is available on our hosted API, with private deployments following soon for select enterprise customers.

Try Thinking on your own data in the playground, and reach out to us if you’d like to be a part of our enterprise offering.

Try Thinking in the playground ↗