LimiX-2 Tops TabArena, Overtaking TabPFN-3.5 in Tabular AI

Two research teams released new tabular AI models within roughly 24 hours of each other in mid-September 2026, and both claimed the same crown. Stable AI and Tsinghua University’s LimiX Team published LimiX-2, a 400-million-parameter foundation model that reports an Elo of 1935 on the TabArena leaderboard, while Prior Labs released TabPFN-3.5, whose strongest variant reports 1910 Elo on the same benchmark.

Tabular foundation models are AI systems built to learn from spreadsheets and databases: the structured data behind fraud detection, credit scoring, medical research, and business forecasting. This article explains what LimiX-2 is, how its scores compare with TabPFN-3.5, why the “number one” claim is more contested than the headlines suggest, and what the licensing fine print means for anyone hoping to use either model in production.

What Is LimiX-2?

LimiX-2 is the newest entry in the LimiX family of large structured-data models. The LimiX Team, credited in the technical report to Stable AI and Tsinghua University, posted the technical report on arXiv on September 15, 2026, then released the model weights and inference code on September 16. Search interest around the release clusters exactly where you would expect: the model family names, the GitHub repository, and the paper itself.

The lineage matters. LimiX V1.0 launched in August 2025 as a pair of small models (16M and later 2M parameters), and a lightweight LimiX-2M variant followed in November 2025.

LimiX-2 is the scaled-up successor: roughly 400 million parameters, trained using scaling laws derived from the first generation, and built to handle classification, regression, and missing-value imputation in a single forward pass, without any task-specific fine-tuning.

Foundation models have expanded well beyond text in recent years, from vision systems to forecasting models such as Google’s TimesFM-3 multivariate time series foundation model. Tabular data arguably remains the largest untouched frontier, since most enterprise data still lives in tables. The official LimiX repository had already gathered more than 4,200 GitHub stars and 300 forks within days of the release, and the LimiX-2 checkpoint is available for download from the Hugging Face model card.

The TabArena Results: Where LimiX-2 Ranks

TabArena is a living benchmark for tabular machine learning: 51 manually curated real-world datasets (30 binary classification, 8 multiclass classification, and 13 regression tasks) scored with Elo ratings anchored on Random Forest at 1000.

It is community-maintained and hosted as a public TabArena leaderboard on Hugging Face, which makes it the closest thing tabular AI has to a canonical ranking.

On the full benchmark, LimiX-2 in its default configuration reports an Elo of 1935, first place across all four predictive metrics in the technical report. The margin is substantial: 117.4 Elo points above the runner-up, TabFM+, before rounding. LimiX-2’s improvability score (3.3% versus 6.2% for TabFM+) and average rank (5.5 versus 9.0) both favor the new model, and its aggregated win count of 18.9 is roughly 3.6 times TabFM+’s 5.3.

In head-to-head terms, the report says LimiX-2 wins about 65% of tasks against TabFM and 67% against EXAONE Tabular.

The subset and cross-benchmark results are equally strong:

• Classification subset (38 TabArena datasets): Elo of 1917 and a 94.5% average pairwise win rate.
• Regression subset (13 TabArena datasets): Elo of 2206, a 96.9% win rate, and improvability of just 0.6%.
• TALENT (288 datasets): 1506 overall Elo, finishing first in all five evaluation categories ahead of TabFM at 1471.
• BCCO (156 datasets, focused on incomplete and challenging data): 1432 Elo, ahead of AutoGluon 1.6 at 1376 and TabFM at 1369.

The pattern holds across all three benchmarks the team evaluated, with the respective runners-up being TabFM+ on TabArena, TabFM on TALENT, and AutoGluon 1.6 on BCCO:

All of these figures are vendor-reported: they come from the LimiX-2 technical report, which states that it followed TabArena’s official evaluation configuration and result-aggregation pipeline. Independent reproduction has not yet appeared, a caveat we return to below.

LimiX-2 vs TabPFN-3.5: A Contested Crown

Here is the detail that most coverage of this leaderboard shuffle misses: Prior Labs released TabPFN-3.5 on September 15, 2026, the same day the LimiX-2 paper appeared, and it too claimed first place on TabArena. The TabPFN-3.5 technical report states that the family ranks first across seven tabular benchmarks, with TabPFN-3.5-Thinking reaching 1910 Elo on TabArena and the base model at 1866, ahead of TabFM+ at 1823 in Prior Labs’ snapshot.

As of this writing, the live TabArena leaderboard reflects LimiX-2 at the top, ahead of TabPFN-3.5’s entries.

Crucially, neither model was benchmarked against the other. The LimiX-2 report uses TabPFN-3, at 1632 Elo, as its Prior Labs baseline, not TabPFN-3.5. The TabPFN-3.5 report, written before LimiX-2’s weights were public, never mentions LimiX at all. The two sets of numbers also come from different snapshots of a living benchmark: Prior Labs records TabFM+ at 1823 Elo, while the LimiX report records the same system at 1818. Comparing 1935 against 1910 is therefore informative, but it is not a controlled head-to-head experiment.

Benchmark crowns in AI move fast. Intelligent Living’s earlier coverage of MiniCPM5-2B topping the Intelligence Index for open-weights models under 4 billion parameters tells the same story: a leaderboard win is a snapshot, not a permanent title. The table below compares what is actually known about the two rival models.

How Contextual Mechanism Networks Work

Most tabular foundation models, including the TabPFN line, are built on Prior-Data Fitted Networks. They are trained to approximate one function: the probability of a target column given a query row and a set of labeled context rows. Supervision within each training task is confined to predicting that single target.

LimiX-2 takes a different organizing principle that the team calls Contextual Mechanism Networks. Instead of learning only the conditional distribution of a designated target, the model learns the joint dependency structure among all variables in a table, conditioned on context.

Under this framing, ordinary supervised prediction becomes a special case of a broader capability: inferring any unobserved value from the available evidence. Pretraining uses Context-Conditional Masked Modeling, which combines target prediction with masked-feature reconstruction across three masking patterns: individual cells, whole columns, and blocks of entries.

The architecture reinforces this design at the representation level:

• Every cell gets its own embedding vector, with dimension 256, rather than compressing an entire row into a single vector.
• A 24-layer stack of dual-axis transformer blocks alternates sample-axis attention, where queries learn from context rows but never from each other, with asymmetric feature-axis attention that lets feature representations read the target pathway but not vice versa.
• Separate gated SwiGLU feed-forward networks process feature representations and task representations independently, with task information carried in four embedding slots per row.
• Regression is handled by partitioning the target range into 5,000 ordered bins and predicting a probability for each, rather than minimizing mean squared error directly.
• Column identity is injected through low-rank discriminative feature encoding, which keeps the model from confusing position with semantics as tables widen.
• Multi-head cross-attention applies key/value normalization plus a learnable length-dependent rescaling, which stabilizes attention behavior as context tables grow.

Training data is entirely synthetic. The team generates datasets from structural causal models: directed acyclic graphs assembled from causal motifs such as chains, confounders, and colliders, with node values produced by sampled functions spanning neural networks, decision trees, CNNs, kernels, periodic and multiplicative interactions, then adapted into classification and regression tasks. The payoff shows up in the efficiency numbers: the report states that LimiX-2 surpasses TabFM despite being roughly four times smaller in parameter count.

Beyond Prediction: Imputation and Causal Skeleton Recovery

A single LimiX-2 checkpoint handles classification, regression, missing-value imputation, and causal discovery without any task-specific parameter updates. That breadth is not just marketing. The BCCO benchmark deliberately stresses incomplete and imperfect datasets, and LimiX-2 reports an Elo of 1432 there, ahead of AutoGluon 1.6 at 1376 and TabFM at 1369, suggesting the imputation-oriented training pays off on messy real-world data.

The causal claims are the most novel part of the release. In causal skeleton recovery evaluations, the report states that attention patterns extracted from LimiX-2 recover direct causal relationships more accurately than other tabular foundation models, than XGBoost-based feature importance, and than dedicated causal discovery methods. For fields like healthcare and economics, where understanding which variables drive outcomes matters as much as predicting them, an attention map that doubles as a causal signal would be a meaningful capability.

Some caution is warranted. These causal evaluations come from the same vendor report as the benchmark scores, the pretraining distribution is synthetic by construction, and causal structure claims carry real-world consequences when they are wrong. Treat the causal-skeleton results as promising research evidence, not as a validated replacement for domain expertise.

The Scaling Story: 400M Parameters Is Not the Ceiling

The most forward-looking finding in the technical report is not a leaderboard score. The team trained LimiX-2 at six sizes, from 12.5M to 406.2M parameters, and found that downstream performance follows a clean log-linear trend with model size across all five evaluation series.

On TabArena, Elo rises from 1766 at 12.5M parameters to 1935 at 406.2M, a fitted gain of 34.68 Elo points per parameter doubling with an R-squared of 0.98. The other series show gains between 11.24 and 30.06 Elo per doubling.

Just as important, the team reports no sign of performance saturation within the measured range, and the report extrapolates the trend toward the billion-parameter regime. If that trajectory holds, today’s 400M-parameter model is an early waypoint rather than a finished product, and the pressure on tuning-heavy automated machine learning pipelines, such as AutoGluon’s four-hour benchmark runs, will keep growing.

The team reports observed Elo of 1766 at 12.5M parameters and 1935 at 406.2M; the fitted line below uses the report’s own scaling coefficients, with asterisks marking extrapolated sizes:

Prior Labs is betting on a different axis with TabPFN-3.5: a 220M-parameter base model, an 84M-parameter Fast variant claimed to run up to six times faster, and support for training sets up to a million rows. Scale versus deployability is shaping up as the central strategic split in tabular AI.

Can You Actually Use LimiX-2?

The engineering story is genuinely open. The LimiX-2 checkpoint (LimiX-2.ckpt) is downloadable from Hugging Face, and the repository ships a pip-installable inference package requiring Python 3.12 or newer with PyTorch 2.9.1. The command-line tool supports classification, regression, and feature imputation tasks, with a device flag that accepts both CPU and CUDA, and a Python predictor class exposes the same three task types programmatically. Within days of release, the repository had collected more than 4,200 stars from a developer community clearly eager to test the top-ranked model.

The licensing picture is more restrictive than the “open-source release” headlines imply. The original LimiX project was announced under the Apache 2.0 license, but LimiX-2 changes the terms: the model weights are governed by the StableAI LimiX Non-Commercial License, and the repository code ships under the Stable AI Technology Co., Ltd. License v1.0, which reproduces Apache 2.0 but adds additional attribution and model-naming requirements.

Research, evaluation, and benchmarking uses appear unrestricted; production deployment terms are not published, so commercial users must negotiate with Stable AI directly.

Prior Labs made the same move with TabPFN-3.5: open weights for research, evaluation, and Kaggle use under a non-commercial license, with production use routed through a paid API or commercial agreement. The pattern is worth noting. The most capable tabular foundation models of 2026 are open in spirit but closed for commercial deployment, which echoes licensing shifts seen elsewhere in AI model releases.

The practical hardware question remains unanswered. Neither the report nor the model card publishes VRAM or memory requirements, inference latency, or throughput figures for the 400M-parameter model. CPU inference is supported in principle, but in-context learning over large tables is memory-intensive, so teams should plan for a GPU and budget time for their own benchmarking before committing workflows to it.

A practical checklist for teams evaluating the model:

• Download the checkpoint from Hugging Face and review the license terms before evaluating production use.
• Install the inference package (Python 3.12 or newer, PyTorch 2.9.1) and run the bundled classification, regression, or imputation examples.
• Benchmark inference on your own datasets and hardware, since published latency and memory figures do not exist yet.
• Compare results against tuned gradient-boosted baselines such as XGBoost, LightGBM, or CatBoost on your data.
• Monitor the live TabArena leaderboard, since rankings on this living benchmark shift as new methods are added.
• Contact Stable AI directly for commercial deployment terms under the non-commercial weights license.

The Catch: Self-Reported Scores and an Unverified Crown

Every headline number in this article comes from vendor technical reports. As of publication, no independent third-party reproduction of either LimiX-2’s or TabPFN-3.5’s TabArena results has appeared, and the two teams’ evaluations were conducted days apart against different baselines.

The Elo confidence interval reported for LimiX-2 itself spans from roughly -77 to +111 points, which is a useful reminder that leaderboard rankings carry statistical uncertainty.

This dynamic is not unique to tabular AI. Benchmark claims across the industry routinely arrive ahead of independent verification, a pattern examined in our analysis of what benchmark wins really mean when a model’s results are self-reported. To their credit, both reports describe credible evaluation protocols: LimiX-2 states that it followed TabArena’s official configuration and aggregation pipeline, and Prior Labs evaluated across all 51 TabArena datasets using the benchmark’s standard task suite.

For readers tracking the leaderboard, the practical guidance is simple:

• Check the live TabArena leaderboard rather than any single announcement.
• Look for independent reproductions before treating any ranking as settled.
• Remember that Elo gaps within confidence intervals are effectively ties.

Frequently Asked Questions

What is LimiX-2 used for?

LimiX-2 is a pretrained foundation model for structured data. It performs classification, regression, and missing-value imputation in a single forward pass using in-context learning, meaning it makes predictions from your dataset without any fine-tuning. Its attention patterns have also been evaluated for causal discovery on tabular data.

Did LimiX-2 actually beat TabPFN-3.5 on TabArena?

On vendor-reported numbers, yes: LimiX-2 reports 1935 Elo against 1910 for TabPFN-3.5-Thinking and 1866 for the base model, and the live TabArena leaderboard reflects LimiX-2 at the top as of September 2026. However, neither team benchmarked the other’s model directly, and the scores come from separate snapshots of a living benchmark, so the comparison is indicative rather than conclusive.

Is LimiX-2 open source?

Partly. The inference code is publicly available on GitHub under a Stable AI license derived from Apache 2.0 with added attribution requirements, and the weights are freely downloadable from Hugging Face. But the weights are governed by the StableAI LimiX Non-Commercial License, so commercial production use requires separate terms from Stable AI. The first-generation LimiX project was announced under Apache 2.0, making LimiX-2’s non-commercial weights license a notable tightening.

Which is better, LimiX-2 or TabPFN-3.5?

There is no verified answer yet. Vendor-reported results favor LimiX-2 on TabArena, TALENT, and BCCO, while the TabPFN-3.5 family offers deployment advantages: a faster lightweight variant, documented support for training sets up to a million rows, and a clearer commercial path through Prior Labs’ API. Both models train exclusively on synthetic data, both ship non-commercial weights licenses, and neither score set has been independently reproduced.

What hardware do I need to run LimiX-2?

The published requirements are modest on paper: Python 3.12 or newer, PyTorch 2.9.1, and either a CPU or GPU. What is missing is specifics. Stable AI has not published VRAM minimums, memory footprints, or inference latency figures for the 400M-parameter model, so anyone planning local deployment should test on their own data first.

What is TabArena?

TabArena is a community-maintained, living benchmark for tabular machine learning hosted on Hugging Face. It evaluates methods on 51 curated real-world datasets spanning binary classification, multiclass classification, and regression, and reports results as Elo ratings anchored on Random Forest at 1000. Because new methods are continuously added, its leaderboard values shift over time.

Conclusion

LimiX-2 arrives with the strongest vendor-reported tabular benchmark results to date: 1935 Elo on TabArena, first place across TALENT and BCCO, and a modeling approach that treats supervised prediction as one case of joint conditional inference rather than the whole objective. Its contextually trained architecture delivers state-of-the-art numbers at a quarter of TabFM’s parameter count, and its scaling results, with no saturation in sight up to 406M parameters, suggest bigger models are coming.

Whether the crown is truly its own remains an open question. TabPFN-3.5 claimed the same benchmark within the same week, the two models have never been compared under one evaluation, and no independent lab has reproduced either set of scores. What is settled is the direction of travel: tabular foundation models are improving fast enough to challenge both heavily tuned AutoML ensembles and gradient-boosted defaults, while their licenses quietly restrict commercial use. The next milestone worth watching is not another leaderboard entry. It is the first independent, apples-to-apples verification of these results.