LimiX

LimiX is a family of foundation models for structured data, meaning rows and columns of tabular data. The current release, LimiX 2, is built around a design the developers call Contextual Mechanism Networks. Instead of training a separate model for each dataset, a single pretrained model reads a table as context and can then answer several kinds of questions about it, including classification, regression, missing value handling and recovery of causal structure. The model is pretrained on synthetic datasets generated from structural causal models with varied graph structures and functional mechanisms. The vendor reports that LimiX 2 ranks first by Elo rating on three tabular benchmarks, TabArena, TALENT and BCCO, ahead of other tabular foundation models and dataset specific systems. It also reports that the feature attention in the model encodes direct causal relationships, which allows recovery of a causal skeleton. The site links to a technical report, a GitHub repository, model pages on HuggingFace and ModelScope, and documentation. It is aimed at data scientists, machine learning engineers and researchers who work with tabular data. The vendor does not publish pricing on its site.