Back to Gap Map Infrastructure / Classic ML & computer vision

LimiX

Stable AI
restricted / Overall score: 2.2

LimiX is a family of foundation models for tabular and other structured data from Stable AI. One pretrained model handles classification, regression and missing-value imputation on a new table in a single forward pass, reading the table's rows as context instead of training on them. The current LimiX-2 has 400M parameters and is pretrained on synthetic data drawn from structural causal models.

Openness

2 high confidence
2.0
weights
open (LimiX-2 checkpoint on the Hub, ungated, beside the earlier LimiX-2M and LimiX-16M)
data
closed (synthetic tables from structural causal models; the generator is not released)
code
partial (inference code only; no pretraining or data-generation code)
license
Stable-AI-License-1.0 (code; Apache-2.0 with added attribution and model-naming terms)+StableAI-LimiX-Non-Commercial-License-1.0 (LimiX-2 weights; no commercial rights)

LimiX-2's weights are under Stable AI's non-commercial license, which rules out production, internal business and hosted use, and the inference code is under an Apache-2.0 variant that adds attribution and model-naming requirements. The synthetic pretraining data cannot be regenerated, since neither its generator nor the training code is released.

Adoption

1 medium confidence
1.0

Hugging Face downloads over the trailing 30 days for the LimiX-2 checkpoint and the two earlier checkpoints the repository links to. Copies on ModelScope are not counted.

Capability

4 medium confidence
4.0

LimiX predicts on an unseen table from its rows in context, with no fitting step, as TabPFN does, and also fills in missing values. Its authors report it ahead of TabFM and AutoGluon on public benchmarks, but it is a single model and does not search over or ensemble other learners.

  • https://arxiv.org/abs/2609.17488 recorded 2026-11-12

    LimiX-2 abstract: "Pretraining uses synthetic datasets generated by structural causal models (SCMs)"; evaluations on TabArena, TALENT and BCCO.

  • https://raw.githubusercontent.com/limix-ldm-ai/LimiX/main/README.md recorded 2026-11-12

    README: "A single pretrained model performs classification, regression and missing-value imputation in one forward pass, without task-specific parameter updates"; "LimiX-2 achieves the highest Elo rating on all three benchmarks, outperforming all compared foundation models and AutoGluon 1.6".

Verified 2026-11-12