LimiX-2

LimiX-2 is a foundation model for structured and tabular data, released by Stable AI with Professor Peng Cui's group at Tsinghua University on 16 September 2026. The first thing a business needs to know about it is the licence: the weights are downloadable, but under "stableai-limix-non-commercial-license-v1.0", which permits non-commercial use only. It is not open source and it is not free to build a product on. What it does is unusual for this board, which is mostly language models: a single pretrained 400M-parameter checkpoint handles classification, regression and missing-value imputation across tables it has never seen, without any task-specific parameter updates — the role a gradient-boosting library and a round of per-dataset training usually play. The architecture is what the authors call a Contextual Mechanism Network, pretrained with Context-Conditional Masked Modeling on synthetic datasets generated from structural causal models, and they report that its feature attention encodes direct causal relationships well enough to recover a causal skeleton, which is a claim about explanation rather than accuracy. Stable AI reports first place on the three tabular leaderboards it was evaluated on, quoting an Elo of 1935 on TabArena (117.4 points above the runner-up), 1506 on TALENT (35 points above TabFM) and 1432 on BCCO (56 points above AutoGluon 1.6). Those are the vendor's own figures from its own model card and paper. They are not stored as benchmark scores here, because they are Elo ratings on tabular leaderboards no other model on this board has been measured against, and an Elo has no meaning next to the percentage scores the rest of the column holds. Inference code and the checkpoint are published on GitHub and Hugging Face.