TLDR: Prior Labs has released TabPFN-2.5, the latest iteration of its tabular foundation model, significantly boosting its capacity to handle larger datasets. This new version supports up to 50,000 samples and 2,000 features, a substantial increase over its predecessor, TabPFNv2. TabPFN-2.5 maintains a training-free workflow and achieves state-of-the-art performance in tabular prediction tasks, making it a powerful tool for industries relying on structured data.
Prior Labs has announced the release of TabPFN-2.5, marking a significant leap forward in the realm of tabular foundation models. This latest version of TabPFN is engineered to unlock unprecedented scale and speed for handling tabular data, which is prevalent in critical sectors such as finance, healthcare, energy, and industry.
TabPFN-2.5 introduces a substantial increase in data handling capabilities, now supporting datasets with up to 50,000 samples and 2,000 features. This represents a five-fold increase in rows and a four-fold increase in columns compared to TabPFNv2, effectively allowing for approximately 20 times more data cells in the supported regime. This enhanced scalability makes TabPFN-2.5 viable for a much wider array of real-world problems.
At its core, TabPFN-2.5 operates on the principle of a prior data-fitted network, leveraging a transformer-based foundation model that utilizes in-context learning to solve tabular prediction problems in a single forward pass. Unlike traditional machine learning approaches, TabPFN-2.5 eliminates the need for dataset-specific gradient descent or hyperparameter tuning, streamlining the workflow and reducing ML engineering overhead.
The model is meta-trained on extensive synthetic distributions of tabular tasks. During inference, users provide training rows and labels along with test rows, and the model directly outputs predictions. This ‘training-free’ workflow is a hallmark of the TabPFN series, offering significant advantages in terms of speed and ease of use.
Performance-wise, TabPFN-2.5 has demonstrated state-of-the-art results across classification and regression tasks. On industry-standard benchmarks like TabArena-lite, TabPFN-2.5 substantially outperforms tuned tree-based models such as XGBoost and CatBoost in a forward pass. It also matches the accuracy of AutoGluon 1.4, a complex ensemble that typically requires four hours of tuning and includes previous TabPFNv2 versions.
Prior Labs has also introduced a new distillation engine with TabPFN-2.5. This engine allows the conversion of TabPFN-2.5 into a compact Multi-Layer Perceptron (MLP) or tree ensemble, preserving most of its accuracy while delivering orders-of-magnitude lower latency. This feature is particularly beneficial for production use cases, enabling plug-and-play deployment.
TabPFN-2.5 is accessible through a Python package and an API, with a non-commercial open-source license available on Hugging Face. An enterprise version, TabPFN Enterprise, is also available on request, offering advanced features like fine-tuning, context reasoning, real-time inference, and support for larger data and causal reasoning.
Also Read:
- New Benchmarking Suite Terminal-Bench 2.0 and Agent Testing Framework Harbor Launched to Advance AI Agent Evaluation
- Anthropic Revolutionizes AI Agent Efficiency with ‘Code Execution with MCP’ Approach
This release builds upon the success of its predecessors: the original TabPFN, which handled up to 1,000 samples and clean numerical features, and TabPFNv2, which extended support to messy real-world data, including categorical features, missing values, and outliers, for up to 10,000 samples and 500 features. TabPFN-2.5 represents the next generation, making prior data-fitted networks even more practical and powerful for real-world tabular problems.
