Prior Labs has released TabPFN-3, a tabular foundation model built for the rows and columns that still drive most business decisions.
TabPFN-3 matters because it takes a category that looked impressive in benchmarks and moves it closer to the operating scale startups and enterprises actually need. Released on May 12, 2026, the model is designed to make predictions on structured datasets with up to 1 million training rows, while keeping inference practical on a single H100 GPU through a reduced KV cache and row-chunked processing.
That sounds technical, but the business point is simple. Most machine learning inside companies is not about writing poems, generating images, or answering support tickets. It is about deciding whether a borrower is risky, whether a machine is likely to fail, whether a patient needs follow-up, or whether demand will move in a certain direction next week. Those problems live in tables.
For years, startups building around that kind of data have leaned on gradient-boosted trees such as XGBoost, LightGBM and CatBoost. They are reliable, well understood and hard to beat. But they also demand feature engineering, tuning, preprocessing and a level of data science discipline that many young companies only build slowly. TabPFN-3 is an attempt to compress that workflow into something closer to a foundation model experience.
As Prior Labs noted in its May 12 technical report, TabPFN-3 builds on earlier releases that moved from small clean datasets to larger, messier business data with categorical features, missing values and outliers. TabPFN-2.5 reached 100,000 rows and 2,000 features. The new release pushes the row count to 1 million, while the company says it is up to 20 times faster than TabPFN-2.5.
That scale matters because a model that only shines on small benchmark tables can be useful in research but awkward in production. Credit scoring, customer churn, fraud detection, clinical risk and maintenance forecasting often do not stop at tidy sample sizes. The difference between 100,000 rows and 1 million rows is the difference between a clever notebook result and something that begins to fit the shape of enterprise data.
Prior Labs is also making a performance claim that will get attention from data science teams. On TabArena, a benchmark for structured data tasks, its API-only TabPFN-3-Plus with Thinking mode beats non-TabPFN models by more than 200 Elo, rising to 420 Elo on the largest-data subset. The company says the standard TabPFN-3 has a 93% win rate over classic machine learning on TabArena, while its product page points to predictions on 1 million samples in 0.2 seconds.
The detail worth watching is Thinking mode. It uses extra inference-time computation to improve predictions, which echoes the direction large language models have taken, but applies it to tables rather than text. For startups, this is the more interesting lesson. The frontier model pattern is spreading into less glamorous parts of enterprise software, and those are often the parts where customers already have budgets.
Licensing turns research into a business
Prior Labs is not treating TabPFN-3 as a pure open research drop. The release is available under the TABPFN-3.0 License v1.0, which the company describes as permissive for research and internal evaluation. Production use, including TabPFN-3-Plus with Thinking mode, is routed through API and enterprise licensing, with deployment options that include on-premise and private cloud environments such as AWS SageMaker and Azure AI Foundry.
That split is important. It lets researchers, builders and internal teams test the model without putting the whole commercial engine in the open. At the same time, it gives Prior Labs a direct path to revenue from companies that want the model inside real workflows. This is where AI infrastructure startups increasingly have to be careful. Adoption is not the same thing as a business model, but a widely used research tool can become one if the production path is clear.
The timing also sits inside a bigger corporate story. SAP announced on May 4 that it had entered a definitive agreement to acquire Prior Labs, with plans to invest more than 1 billion euros over four years to scale it into a frontier AI lab for structured business data. SAP said the transaction is expected to close in the second or third quarter of 2026, subject to regulatory approval, and that Prior Labs will continue to operate independently.
That gives TabPFN-3 a different context than a normal model release. It is not just a lab showing a new benchmark. It is a startup, now on a path toward a major enterprise software owner, trying to define a category around tabular foundation models. SAP has obvious reasons to care. Its customers run on ledgers, inventory tables, procurement records, payroll systems and customer databases. If AI is going to move from chat interfaces into operational decision-making, those rows and columns have to become first-class model inputs.
For founders, the takeaway is not that every startup should replace its current machine learning stack tomorrow. The release still needs independent testing in messy production environments, and licensing will matter for anyone planning to build a product on top of it. But the direction is clear. Foundation models are moving into the business data layer, where results are judged by accuracy, latency, deployment control and whether the model improves decisions people already pay to make.
The next thing to watch is whether TabPFN-3 changes buying behavior, not just benchmark tables. If data teams can get strong predictions with less tuning and fewer brittle pipelines, a new class of AI infrastructure will start to look less like research tooling and more like the default way companies model their own operations.
Also read: Japan's robot lab shows physical AI is moving into drug research • eBay just forced Ryan Cohen to prove GameStop can pay. • A cup of water exposed the real economics of local AI hardware
