LimiX-2: 400M Tabular Model Handles Classification, Regression and Imputation

By AiCybr Editorial · Published September 20, 2026

Stable AI and Tsinghua University collaborators released LimiX-2 on September 16, 2026, with a downloadable 400-million-parameter checkpoint and inference code for structured data. One pretrained model performs classification, regression and missing-value imputation in a forward pass without task-specific parameter updates.

The release targets the tabular workloads behind databases, spreadsheets and analytical pipelines. Its Contextual Mechanism Network architecture learns a context-dependent joint representation of features and targets, while pretraining uses synthetic datasets generated from structural causal models. The team reports first-place overall Elo ratings on the TabArena, TALENT and BCCO benchmark suites.

LimiX-2 is available for local inference, with one important deployment constraint: the released model is governed by the StableAI LimiX Non-Commercial License. Commercial teams therefore need a separate licensing decision before adopting the checkpoint in production.

LimiX-2 specifications

The CMN design changes the modeling target used by many prior-fitted tabular networks. LimiX-2 models the joint structure of features and targets conditioned on the reference dataset. That design lets the same checkpoint answer different structured-data questions from in-context examples.

Benchmark results reported by the LimiX team

The team's published evaluation reports 1935 overall Elo on TabArena, 1506 on TALENT and 1432 on BCCO. In its TabArena comparison, LimiX-2 leads the runner-up TabFM+ by about 117 Elo points. The TALENT results place it first across the reported overall, classification, regression, binary-classification and multiclass-classification categories. BCCO adds tests involving challenging and incomplete data.

These are results published by the LimiX team and should be evaluated as author-reported benchmark evidence. The public weights make independent reproduction possible, and production selection should add application-specific validation against the incumbent tabular stack.

A particularly useful comparison is against mature dataset-specific systems such as AutoGluon, CatBoost, XGBoost and LightGBM, plus other tabular foundation models. Conventional AutoML systems can train or tune models for each dataset; LimiX-2's value proposition is a reusable pretrained checkpoint that can make predictions from the supplied context without per-dataset parameter training.

Scaling from 12.5M to about 406M parameters

The accompanying work evaluates six model sizes from 12.5 million to 406.2 million parameters. The team reports approximately log-linear downstream performance trends across its evaluated classification and regression series, with no clear saturation within that tested range.

That result is relevant to the direction of tabular foundation models because it provides an empirical case for scaling model capacity beyond the small checkpoints common in this category. It remains a measured trend over the studied range; larger billion-parameter variants will need their own efficiency and quality measurements.

The released model is described publicly as 400M parameters, while the scaling study gives 406.2M for its largest configuration. Those figures describe the same approximate model scale and are best treated as marketed and technical precision levels, respectively.

Classification, regression and imputation in one checkpoint

A conventional structured-data pipeline often separates classification, regression and imputation into different estimators or processing stages. LimiX-2 exposes all three tasks through one pretrained model.

For classification, the model consumes a reference table and produces class probabilities for query rows. Regression produces numerical predictions. Missing-value imputation uses observed features and context to fill missing entries. The model's documented workflow performs these tasks without updating checkpoint weights for each new dataset.

This can simplify experimentation where teams maintain many heterogeneous tables, especially when the cost of repeated model selection and hyperparameter tuning is significant. The practical comparison should include accuracy, inference latency, accelerator requirements, dataset-size limits and operational complexity against the existing pipeline.

Synthetic causal pretraining

LimiX-2 is pretrained on synthetic tables generated using structural causal models spanning different graph structures, functions and observation processes. This approach lets the training generator produce large numbers of structured-data problems with controlled relationships between variables.

The team also reports that feature attention contains information useful for recovering causal skeletons. That is a research result about the learned representation; predictive deployment and causal inference remain separate validation problems, and high predictive benchmark scores alone do not establish causal effects in a production dataset.

Local deployment and licensing

The released checkpoint and inference code make LimiX-2 directly testable on local infrastructure. The project's documentation provides the model artifact through its public distribution channels and describes CPU/GPU inference workflows.

Licensing is the main immediate production constraint. LimiX-2 uses the StableAI LimiX Non-Commercial License v1.0, so the checkpoint should not be treated as a permissively licensed commercial model. Teams evaluating it for revenue-generating products or internal commercial deployment should review the license terms and obtain appropriate permission where required.

This differs from the earlier LimiX project's Apache-2.0 release history. License checks should therefore follow the specific LimiX-2 artifact instead of inheriting assumptions from previous family members.

Where LimiX-2 fits

LimiX-2 is most relevant to teams evaluating pretrained models as an alternative or complement to per-dataset tabular training. Its strongest published case combines one reusable checkpoint, three task families and leading author-reported results across three broad benchmark suites.

The next decision is operational. Benchmark it against the current AutoML or gradient-boosting pipeline on representative tables, including categorical-heavy data, missing values, regression targets and the dataset sizes expected in production. Record quality, latency, memory use and total pipeline time under the same train/test splits.

Bottom line

LimiX-2 moves the LimiX family to a 400M-class structured-data foundation model and packages classification, regression and imputation into one released checkpoint. Stable AI and Tsinghua's team reports overall Elo scores of 1935 on TabArena, 1506 on TALENT and 1432 on BCCO, with weights and inference code available for independent testing.

The combination is technically significant for tabular AI, while the non-commercial model license materially limits straightforward production adoption. For research and evaluation, the public checkpoint provides a concrete way to test whether pretrained tabular inference can replace part of a conventional dataset-specific training workflow.

Sources

• LimiX-2 model card, release details and benchmark tables: https://huggingface.co/stable-ai/LimiX-2
• LimiX-2 technical report: https://arxiv.org/abs/2609.17488
• StableAI LimiX Non-Commercial License v1.0 for LimiX-2 weights: https://huggingface.co/stable-ai/LimiX-2/blob/main/LICENSE