Tabular data is the most common data form in enterprises, yet the most neglected by the foundation-model wave. On September 16, Stable AI released the weights and inference code of LimiX-2 to its GitHub repo: a 400M-parameter pretrained model that handles classification, regression, and missing-value imputation in a single forward pass, with no task-specific parameter updates. The paper landed on arXiv:2609.17488 a day earlier and reached #1 on Hugging Face Daily Papers that day.
First on all three benchmarks — against real opponents
The official scores are the hardest part of this release. On TabArena, LimiX-2 scores an overall Elo of 1935, leading runner-up TabFM+ by 117.4 points with an aggregated win count of 18.9 — roughly 3.6x the opponent's. On TALENT it reaches 1506, 35 points ahead, ranking first across all five evaluation categories. On BCCO it hits 1432, leading AutoGluon 1.6 and TabFM by 56 and 63 points respectively. The TabArena classification subset (38 datasets) shows a 94.5% win rate; the regression subset (13 datasets) shows 96.9%.
The comparison field is not padding: AutoGluon 1.6 is Amazon's AutoML system; TabFM, TabPFN-3, EXAONE Tabular, and Mitra-v2 are the main players of the tabular foundation-model lane; CatBoost and RandomForest are the floor of the gradient-boosting camp. A single 400M pretrained model puts all of them behind — on TabArena, TabPFN-3 trails by over 300 Elo points.
From "predict the target column" to "learn the mechanism"
The core change is what gets modeled. Most tabular foundation models organize around p(y|x, D_context) — given features, predict the target column — which is still "answering a designated question." LimiX-2 adopts the CMN (Contextual Mechanism Network) paradigm proposed in the paper, shifting to learning p(x, y | D_context): the joint generative structure of the whole table, from which any variable can be inferred from the rest.
In pretraining this becomes CCMM (Context-Conditional Masked Modeling): the same table is masked under different observation patterns, turning one table into many conditional prediction problems — inferring different unknowns from different evidence. Training data comes entirely from synthetic datasets generated by structural causal models (SCMs), spanning diverse graph structures, functional mechanisms, and observation processes, with no real tabular corpora involved. A byproduct is causal awareness: the model's feature attention encodes direct causal relationships and can be used for causal skeleton recovery — beyond prediction, into understanding how variables influence each other.
400M is far from the ceiling
The scaling section may be the signal most worth watching. The team evaluated five parameter series from 12.5M to 406.2M; downstream performance follows a clear log-linear trend with model size, with no sign of saturation within the measured range. On TabArena, the fitted trend gives roughly 34.7 Elo per parameter doubling, with R² at 0.98. The curve has room left — bigger tabular foundation models still buy steady gains.
Engineering openness is solid: inference ships as a pip-installable package (Python 3.12 + PyTorch 2.9.1, CPU and GPU), the LimiX-2 checkpoint is on Hugging Face, and the GitHub repo gathered 4.2k stars and 304 forks within days.
Cold water: license and self-reported scores
Two caveats. First, the weights ship under the StableAI LimiX Non-Commercial License. The repo code license is Apache-2.0-based but adds attribution and model-naming provisions. Teams planning production deployment should read the terms first. Second, all scores above are vendor self-reported; no independent third-party reproduction has appeared yet, so "first on everything" currently holds only as an official claim. Community comments so far are congratulatory — substantive challenges or verifications have not shown up.
For the industry, the significance is the route itself: after language models proved "pretrain + no fine-tuning," this is the first time a team has used scaling laws to push the same paradigm to first place across three tabular benchmarks, with an explicit unsaturated extrapolation ahead. The heaviest enterprise workloads were never poem-writing — they were tables. If this lane keeps paying off log-linearly, the default status of AutoML's four-hour tuning sessions may not hold.
Refs: arXiv:2609.17488 / GitHub: limix-ldm-ai/LimiX / HF model page