LimiX-2

A 400M-parameter structured-data foundation model from Tsinghua's Cui Peng group and the startup Stable AI, and the most-upvoted paper of its week on HuggingFace. LimiX-2 replaces the target-centric objective of tabular PFNs with a Contextual Mechanism Network that models the joint structure of the data-generating process, pretrained with Context-Conditional Masked Modeling on datasets synthesised from structural causal models and scaled by the family's earlier scaling laws. One model handles classification, regression, and imputation in a single pass with no task-specific updates; the report puts it first on TabArena (Elo 1935), TALENT (1506), and BCCO (1432). Weights are released under the StableAI LimiX Non-Commercial License; the repository has over 4,000 stars.