Tabular artificial intelligence startup Prior Labs GmbH today announced a new foundation model that can handle millions of rows of data to give enterprises a way to understand and use their most complex, business-critical information.

Described as a first, the new TabPFN model marks a 1,000-fold leap in dataset size in less than a year and brings production-grade accuracy to enterprises at scale.

Tabular data – the structured rows and columns that run the majority of enterprise systems — forms the backbone of financial records, supply chains and customer databases, but the company says progress in tabular AI has lagged fields such as vision and language because the datasets share few consistent patterns.

Unlike images or text, tabular datasets have their own structure and behavior, meaning that models can’t rely on the same underlying signals when learning from healthcare records as they do from financial transactions.

Prior Labs’ new TabPFN model is purpose-built for tabular data. The model is trained on hundreds of millions of synthetic datasets to achieve accuracy without task-specific training, instantly learning patterns from any dataset.

The new model follows a year of impressive development. Prior Labs started the year with a TabPFN model that supported 10,000 rows in January, to 100,000 rows in early November and now today, it debuts a model with support for 10 million data points.

As a foundation model, TabPFN can also be fine-tuned on a company’s own data to create a self-reinforcing cycle of scale. The result, the company says, allows enterprises that adopt TabPFN for their most complex data challenges to sharpen the model’s accuracy and performance.

The company’s models are already finding willing customers, with Hitachi Ltd. using TabPFN for predictive maintenance across its rail network, identifying track issues earlier and reducing manual inspections. U.K.-based biotech Oxford Cancer Analytics is using TabPFN to enable detection of complex lung diseases to enable better patient outcomes, demonstrating the model’s versatility.

“Our heavy focus on world-leading research has really paid off,” said Frank Hutter, chief executive officer and co-founder of Prior Labs. “Scaling to millions of data points in less than a year… is a testament to the speed at which our exceptional team can execute research and put it into production.”

Prior Labs is a venture capital-backed startup that has raised €9 million ($9.34 million) in seed funding. Investors in the company include Balderton Capital Management Ltd., Hector Foundation, Atlantic Labs GmbH, Galion.exe and XTX Markets Ltd.

Image: SiliconANGLE/Ideogram