Prior Labs has released TabPFN-3, a new foundation model built specifically for tabular data. Its headline feature: out-of-the-box, no tuning required. Instead of spending hours writing code and tweaking hyperparameters when you get a new table, you just feed it to the model and get predictions after a single inference pass. The new version pushes the limit to 1 million rows × 200 features, all runnable on a single H100 GPU.
On the standard TabArena benchmark, TabPFN-3 with a single inference beat every traditional tree model that had been meticulously tuned. The company also launched TabPFN-3-Plus, an API version with a "thinking mode" that completes its run in less than a tenth of the time AutoGluon 1.5 takes, while scoring significantly higher on predictions.
The new model is up to 20x faster than its predecessor and now supports multi-class classification, relational data, and time series forecasting.
Licensing: TabPFN-3 weights are free for research and internal evaluation only. Commercial decision-making and customer delivery require a separate license. For enterprises, the real takeaway is that structured data prediction has finally moved from "painstakingly tuning XGBoost" to something as simple as calling a large language model.