The Pulse
Prior Labs releases TabPFN-3.5 for harder tabular prediction
Prior Labs has released TabPFN-3.5, a tabular foundation model designed for messy, high-cardinality, temporal and text-rich data. The company says the model leads TabArena and BeyondArena, while a faster alpha version targets latency-sensit
AI.info Team ·
“Zero-shot inference surpassing my carefully tuned models, consistently taking the top rank in my backtests. TabPFN proved we could move to an inference-only foundation model.”
Jean Pablo Fuquen, AI Engineer, Scaleway
TabPFN-3.5 targets the data that conventional models dislike
Prior Labs released TabPFN-3.5 on September 15, positioning the model as a foundation system for prediction on structured data that does not fit the assumptions of standard machine-learning pipelines. The release covers classification and regression workloads involving mixed data types, high-cardinality fields, temporal relationships, grouped records and text embedded in tables.
Prior Labs says the model is designed for situations in which rows are not independent and columns are not cleanly encoded. Its examples include insurance records that combine claim amounts with reviewer conclusions, transaction data connected to merchants and customers, and industrial or scientific tables containing hundreds of sensor measurements collected across locations or time periods.
Those workloads have traditionally required preprocessing, feature engineering and repeated model tuning. TabPFN-3.5 instead uses in-context learning to make predictions from a dataset in a forward pass, building on the TabPFN approach of moving much of the model-building work into pretraining.
51 TabArena datasets and 142 BeyondArena datasets
Prior Labs reports that TabPFN-3.5 ranks first on both TabArena and BeyondArena. TabArena is an IID benchmark containing 51 curated datasets and more than 27 methods, including more than 10 tabular foundation models. BeyondArena covers 142 datasets with high-dimensional, grouped, temporal, high-cardinality and text-rich data.
On TabArena, Prior Labs says TabPFN-3.5 outperforms the previous leader. On BeyondArena, the company reports an advantage of about 150 Elo points over the previous overall leader. On the benchmark’s non-large datasets, the model leads on text-rich, high-cardinality and high-dimensional data, with an advantage of up to about 250 Elo points over the strongest previous baseline. Prior Labs says it matches that baseline on grouped data.
The results are company-reported benchmark claims rather than an independent evaluation. The technical report links to the benchmark repositories and describes the evaluation protocol, datasets, model configurations, runtime environment and full results.
Plus and Thinking extend the same release
Prior Labs is updating two existing variants alongside the base model. TabPFN-3.5-Plus adds processing for text-rich tables containing product descriptions, customer reviews and internal notes. The company says Plus ranks second behind TabPFN-3.5-Thinking on STRABLE, a benchmark covering 108 tabular datasets with messy strings.
Thinking is aimed at cases where additional compute is acceptable in exchange for higher scores. Prior Labs reports that TabPFN-3.5-Thinking adds 44 Elo points on TabArena and about 20 Elo points on BeyondArena compared with the base TabPFN-3.5 model. The company also says the variant improves results on temporal and grouped data, such as forecasting sales at new stores or predicting future sales from historical records.
“On par with my ML model built over a week, in a couple of clicks,” said Corentin Garet of Beside in a customer comment carried by Prior Labs. Kacper Wieczorek, lead data scientist at Marshmallow, said TabPFN had outperformed gradient boosting on real-world insurance datasets without tuning and produced prediction confidence intervals useful for risk applications.
TabPFN-3.5-Fast arrives in alpha
The release also introduces TabPFN-3.5-Fast, an alpha model for latency-sensitive applications. Prior Labs says it runs up to six times faster than the base model. The company’s product page lists inference latency of 0.17 seconds for TabPFN-3.5-Fast, 0.5 seconds for TabPFN-3.5 and 7 seconds for other tabular foundation models when processing 1,000 rows without a cache.
Prior Labs says the new family can handle datasets with up to 1 million rows natively. Its product materials also claim that TabPFN-3.5 is 20 times faster than earlier model versions, with speedups of up to 1,000 times at scale. Those figures describe the company’s own configurations and comparisons, so deployment hardware and workload shape the practical result.
API access arrives with a limited discount
The TabPFN-3.5 family is available through Prior Labs’ API, which ships Plus, Fast and Thinking variants. The company’s MCP service now uses TabPFN-3.5-Plus. API and MCP users receive a 50% reduction from the standard TabPFN-3.5 token rates between September 15 and September 29, 2026; standard rates resume on September 29.
SAP customers can access TabPFN-3.5-Plus through SAP AI Core, while Plus and Thinking are available on AWS SageMaker. Prior Labs says Microsoft Azure ML support will follow. The open-source tabpfn package ships the base model and Fast checkpoint.
The downloadable model card lists three checkpoints, including the standard model, the smaller Fast version and an experimental multiclass variant. It says TabPFN-3.5 is trained purely on synthetic tabular tasks and supports classification and regression on structured data, including string and datetime columns.
That release is not an unrestricted commercial deployment. The model weights use the TabPFN-3.5 license, which permits research, testing and limited internal evaluation but bars commercial or production use of the model, its derivatives and its outputs. Prior Labs directs production users toward a separate commercial enterprise license.