LimiX

Why we included this project

LimiX targets teams that live in spreadsheets and relational tables, where most real-world data actually sits. Instead of building a separate model for every prediction task, it learns a single joint distribution over columns and missing values, so the same checkpoint handles classification, regression, imputation, and tabular generation as different queries. That collapses the usual workflow of training and tuning a fresh model per dataset. The project releases two sizes, LimiX-16M and the lighter LimiX-2M, which uses less GPU memory and runs faster, all under Apache 2.0. The repo's benchmark comparisons against XGBoost and older tabular deep learning approaches across several mainstream datasets give you a concrete sense of where it stands before you invest time.