Skip to main content
The Thinking upgrade is live, and applications are open for our Academic Plan and Research Compute Grants. This update covers everything we’ve shipped since September: tabpfn 9.2.0, tabpfn-client 0.7.1, the Thinking upgrade and tabpfn-extensions 0.6.3.

​
tabpfn-client and API

​
New Thinking mode pushes the Pareto frontier

TabPFN-3.5-Thinking is our new mode of choice for large data. It now supports 1 million training rows, up from 200,000. On BeyondArena-large, covering datasets up to 1 million training rows, Thinking achieved a 95.2% win rate against RealMLP and 90.8% against Kumo-Tabular. We also optimized the speed — new Thinking took 32% less time to predict than the previous version on large-data tasks. See the benchmark results for more details. Now, six effort options alongside the default medium give you fit budgets from 5 minutes to 16 hours. Start with mini, which allows up to 15 minutes for fitting. If the fit finishes successfully before the limit, increasing the effort is unlikely to improve the result. The API reserves the maximum fit charge upfront, then charges for the actual compute spent, subject to a minimum charge. Unused tokens are returned. Check the cost estimate before starting a long fit. Use tabpfn-client 0.7.1 or later:
Read the Thinking guide for fit budgets and examples.

​
Academic Plan and Research Compute Grants

Students and academics can apply for twice the regular monthly API allowance. Verify your academic email and provide your institution and current role. You can keep your existing login email. Need more compute for a research project? Apply for a Research Compute Grant with a short description of the project, the compute you need, your timeline and planned public outputs. We review applications individually. You can apply in account settings..

​
Other API improvements

  • Resume long jobs. Start a fit or prediction, save it, and return to the result from another Python session. See how to run asynchronous fits and predictions.
  • SHAP explanations use fewer tokens. Explain multiple rows together with the client’s SHAP helper. Try the client example.
  • Less overhead for small predictions. Small test sets can travel inside the prediction request, avoiding a separate upload. The client also reuses idle connections for longer.
  • Choose how text is processed. Both advanced (the default) and simple support text columns. Compare the presets.
  • Clearer failures and safer retries. Job errors distinguish failed fits, failed predictions and models that need refitting. The client avoids automatically resending requests after ambiguous failures that could run and bill twice.
  • Pandas period columns work in fit and predict, including year columns.

​
Open-source package

Use tabpfn 9.2.0 or later for the improvements and fixes below. Cached fits use less memory, and fitted models take less space to save. Optional PyTorch compilation can reduce overhead for repeated local predictions. We also fixed:
  • Predictions beyond the training range. TabPFN-3.5 no longer extrapolates too far past a feature’s range, including repeated time values such as one year per monthly row.
  • Regression on tightly clustered targets. We improved predictions when target values have a large mean but little variation. Quantile and tail-distribution calculations are more accurate too.
  • Target proportions after subsampling. Probabilities and regression distributions account for the change in target proportions when subsampling the majority value.
  • Constant or missing columns. Columns that were constant or entirely missing during fitting no longer cause a string-conversion error at prediction time.
TabPFN’s code is now licensed under Apache 2.0; model weights retain their own license.

​
Extensions and relational predictions

TabPFN-Rel now has a standalone package for prediction tasks across related tables. Its agent skill helps Claude Code or Codex prepare a database, define a prediction task and compare results on held-out data. We also released two brand-new extensions:
  • Zero-inflated data. AutoHurdleRegressor supports mean, median and quantile predictions.
  • Images alongside tabular features. TabPFNWithImages accepts image columns. Install tabpfn-extensions[image] to use it. Get started with image features.

​
Migrating

​
API

  1. Install tabpfn-client 0.7.1 or later before using the Thinking upgrade. Version 0.7.0 will be deprecated on 7 January 2027. Upgrade before that date.
  2. Fit a new model. Saved Thinking fits from before the upgrade will no longer work.
  3. Leave thinking_timeout_s unset to use fit budgets above 40 minutes. An explicit timeout is limited to 2,400 seconds and can shorten the selected budget.

​
OSS (local / Hugging Face)

  1. Rerun validation after upgrading: the prediction fixes above can change results.
  2. With majority_downsample, score datasets individually. Batched prediction methods do not support this sampler.
  3. If you construct local models in memory, replace ClassifierModelSpecs and RegressorModelSpecs with ModelSpecs.
Browse the Python and REST reference for signatures and request formats.