Skip to main content
This update covers everything we’ve shipped since May 2026: tabpfn 8.4.0, tabpfn-client 0.4.2, the latest shipped TabPFN API release, and tabpfn-extensions 0.6.0. During this time, we’ve also introduced TabPFN Cookbooks and Research Applications, sharing how TabPFN can be applied for specific use cases. Check them out—and contribute your own.

​
Open-source package

tabpfn 8.4.0 ships with improvements across regression, fitting performance, and Apple Silicon inference:
  • Better uncertainty estimates for regression. Made for forecasting, risk estimation, and other workflows where interval quality matters alongside point predictions. TabPFNRegressor can now calibrate its ensemble distribution against held-out data using NLL or CRPS, improving probabilistic outputs and predicted quantiles (#1172).
  • Faster, lower-memory fitting on large datasets. Made for large or wide datasets—we’ve cut transient memory use and fit time by roughly two-thirds in benchmarks on a 5.3 GB numeric table. The data-cleaning path now avoids redundant float64 copies, assembles encoded arrays in place, and performs pandas dtype casting more efficiently (#1173).
  • Local inference on Apple Silicon. When running TabPFN on a Mac, you can now use MPS inference with MLX-backed FlashAttention and autocast, along with fixes for out-of-memory recovery and incorrect linear-layer outputs.
We also shipped:
  • n_estimators now defaults to "auto" (#1171).
  • Calibrated regression now fails explicitly for batched predictions instead of returning uncalibrated results (#1172).
  • DDP fine-tuning now distributes estimator activations across available GPUs. Read the fine-tuning docs.
  • Regression now supports batched inference (#1164).

​
RelArena-α, TabPFN-Rel, and RPI

This week, we released RelArena-α alongside a family of open-source tools for prediction tasks over relational data. The full benchmarking framework and accompanying tools are described in the technical report. Read why we’re focusing on relational.
  • Compare relational models under a shared evaluation protocol and run benchmarked methods on your own relational data. RelArena-α standardizes data loading, tuning, and evaluation across relational learning tasks, making results easier to reproduce and compare. The Relational Predictive Interface (RPI) provides a common interface for applying these methods to real-world prediction tasks.
  • Apply TabPFN to relational prediction tasks. TabPFN-Rel is a relational harness for TabPFN-3 and TabPFN-3-Plus, and is currently the No. 1 model submission on RelArena-α. Get started with the TabPFN-Rel cookbook.

​
tabpfn-client and API

tabpfn-client 0.4.2 and the TabPFN API ship with improvements across repeated inference and integrations:
  • Up to 10× faster repeat inference with KV cache. Best for large prediction volumes against the same training data and long-running interpretability workloads. fit_with_cache can now be requested during fit() through tabpfn-client 0.4.2 and the hosted TabPFN API, reusing the training-set computation so subsequent predictions avoid the full forward pass and can run in under a second. Get started with KV cache.
We also shipped:
  • TabPFN-3-Plus, including Thinking mode, can now run on Azure AI Foundry through a new estimator interface. Get started with Azure AI Foundry.
  • output_type="full" now returns a consistent response type.
  • Asynchronous fitting is now available through tabpfn-client 0.4.2 and the hosted TabPFN API, improving reliability. See the API reference.
  • The MCP integration now supports asynchronous operations and the standard authentication path. Read the MCP docs.

​
Extensions

tabpfn-extensions 0.6.0 ships with improvements across interpretability and client compatibility:
  • See which training rows drive a prediction. Best for explaining borderline multiclass predictions, investigating unexpected outputs, and seeing how the training set affects results. The same readout can be viewed in target-conditioned embedding space or in the original feature space (#1142).
Two decoder-head attention readouts over target-conditioned embeddings: a confident negative prediction and a borderline negative prediction
We also shipped:
  • The client-backed regressor now preserves output_type="full" and forwards requested output types correctly (#354).

​
Migrating

Review these behavior changes, deprecations, and removals before upgrading production code.

​
OSS (local / Hugging Face)

  1. Review n_estimators when upgrading: it now defaults to "auto", feature-coverage scaling applies only to "auto", and an explicit integer is used exactly as provided. Replace auto_scale_n_estimators, which is deprecated and scheduled for removal in the next major release.
  2. If you calibrate regression with tuning_config and need to score a batch of datasets, call predict() on each dataset individually: predict_batched does not support calibrated regression and now raises NotImplementedError instead of returning uncalibrated fused-batch predictions.
  3. Use PyTorch 2.6.0 or newer for MPS inference.

​
Extensions

  1. Include every feature in the dag mapping passed to TabPFNUnsupervisedModel.generate_synthetic_data() or impute(); declare parentless features with an empty list.

​
API

  1. Reinstall the MCP integration to pick up the latest changes.