The Prior Changelog: Thinking edition
Last time, I promised we’d have to earn your attention through new releases. They are here - Thinking now handles a million rows, students and academics can apply for more tokens, and explaining predictions is more efficient. There’s also more - so make sure to read until the end.
New Thinking mode defines SOTA
Starting today, the new TabPFN-3.5-Thinking is available on our API. It achieved a 95.2% win rate against RealMLP and 90.8% against Kumo-Tabular on BeyondArena-large. See the benchmarks.
This is possible because Thinking now supports 1 million fitting rows, up from 200,000. We recommend trying it on larger datasets, and playing around with brand-new effort levels.
Working on a thesis, paper or research idea?
Our brand-new Academic Plan gives students and academics the extra credits you need to make it happen. For projects that need more, we’re also accepting Research Grant applications. Tell us what you’re working on and how many tokens you need - we review each application individually. Apply now →
New extensions: image and zero-inflated data support
If you are working with multimodal data, or heavily skewed target distributions, these two new extensions are for you:
- Zero-inflated data. For targets with lots of zeros,
AutoHurdleRegressorimproves the results drastically. - Images alongside tabular features. Put product photos, diagnostic scans or other images next to the rest of your table with
TabPFNWithImages. Get started with the image cookbook
Run long jobs and come back later
You no longer need to keep a Python session open while a fit or prediction runs on our API. Submit the job, save the reference, and pick up the result from another session. This improvement is only available on tabpfn-client 0.7.1 and up. We’ll be deprecating the 0.7.0 client version on Jan 7th - so make sure to upgrade.
We also released a standalone TabPFN-Rel with agent skills, local memory use improvements, efficient interpretability on the API and more.
