Stable AI introduced the LimiX-2 model for working with structured data

Stable AI company released LimiX-2 — a specialized fundamental model with 400 million parameters, designed for processing tabular data. The system architecture uses a contextual network mechanism to perform classification, regression, and data imputation tasks in one pass.

The model demonstrates high computational efficiency, achieving a result of 11 points with a training time of just 1 hour. The developers focused on optimizing work with incomplete or noisy datasets, which allows the model to effectively restore missing values in automatic mode.

During testing, LimiX-2 showed superiority over existing analogs in specialized benchmarks TabArena and TALENT. The compact size of the model allows it to be deployed on standard equipment without the need for clusters with high-density GPU.

Integrating the model into data analysis pipelines reduces the time spent preparing datasets for machine learning. Using a single mechanism for diverse classification and regression tasks simplifies the architecture of analytical systems, reducing the number of necessary information preprocessing stages.