Stable AI · 2026-09-16 · major
LimiX-2 — a 400M model tops three structured-data benchmarks
LimiX-2 is a 400M-parameter foundation model for structured data from Stable AI. A single pretrained model does classification, regression and missing-value imputation, and it leads the TabArena, TALENT and BCCO benchmarks.
One 400M model handles classification, regression and missing values on tables, with no per-dataset training.
Quick facts
| Maker | Stable AI |
|---|---|
| Parameters | 400M |
| Tasks | Classification, regression, imputation |
| Pretraining | Context-Conditional Masked Modeling |
| TabArena Elo | 1935 (TabFM+ 1818) |
| License | StableAI LimiX Non-Commercial v1.0 |
| Released | 16 September 2026 |
What is it?
LimiX-2 drops the usual step of training a fresh model for every table. Stable AI released the 400M-parameter model on 16 September 2026 with weights and inference code. One pretrained network answers classification, regression and missing-value questions in a single forward pass, so a new dataset needs no task-specific parameter updates.
How does it work?
The design follows what the team calls Contextual Mechanism Networks. Earlier tabular models centre on p(y|x, D_context), the target-centred objective. LimiX-2 instead learns p(x, y|D_context), a context-dependent view of the joint structure behind how the data was generated. Pretraining uses Context-Conditional Masked Modeling on synthetic tables built by structural causal models that span many graph shapes, functional mechanisms and observation processes.
Why does it matter?
Most business data still lives in tables, and the usual answer is a hand-tuned gradient-boosting model for each one. Stable AI reports that a single pretrained network beats that per-dataset work on all three public suites it tested. The model's feature attention also encodes direct causal relationships, so it can recover a causal skeleton — something a boosted-tree baseline does not give you. The licence is non-commercial, so this is for research and internal evaluation rather than a shipped product.
Who is it for?
data scientists working with tabular data
Frequently asked questions
- Can I use LimiX-2 in a commercial product?
- LimiX-2 ships under the StableAI LimiX Non-Commercial License v1.0, so commercial use is not permitted. Research, benchmarking and internal evaluation are allowed. Teams that need a tabular model inside a paid product should stay on a permissively licensed option such as AutoGluon, which LimiX-2 benchmarks against directly.
- How does LimiX-2 compare with AutoGluon 1.6?
- On the BCCO benchmark LimiX-2 scores 1432 Elo against 1376 for AutoGluon 1.6. The two work very differently: AutoGluon builds and stacks many models for each dataset you give it, while LimiX-2 answers from one pretrained network in a single forward pass. The comparison is a per-dataset pipeline against a general model.
- What do I need to run LimiX-2?
- LimiX-2 needs Python 3.12 or newer and torch 2.9.1, with optional flash-attn support. The published example runs on a CUDA device. You clone the LimiX repository, install it in editable mode, download the checkpoint from Hugging Face, then point LimiXPredictor at the .ckpt file together with an inference config file from the repo.
- How big is LimiX-2 next to a typical language model?
- LimiX-2 has 400M parameters, which is small beside a modern LLM and light enough for a single GPU. Structured-data modelling does not need language-scale capacity: the model learns the mechanisms behind tables rather than world knowledge, so the effort goes into the pretraining objective and the synthetic causal data instead of raw parameter count.
Try it
git clone https://github.com/limix-ldm-ai/LimiX.git && python -m pip install -e .