Wenzhun Intelligent and Tsinghua Release LimiX-2, Structured-Data Foundation Model Tops International Benchmarks

Elo is a ranking metric for measuring a model’s overall capability. TabArena, BCCO and TALENT are mainstream international benchmarks for structured-data models, covering multiple task types and used to assess generalization and predictive performance across different datasets. On TabArena, LimiX-2’s Elo scores for regression, binary classification and multiclass classification were all the best, QbitAI reported.

The upgrade follows three technical routes. The first is Contextual Mechanism Networks, or CMNs, a system for predicting dependency relationships that moves modeling focus from a pre-specified target variable to associations among variables. The second is an upgraded automated synthetic-data generation engine used during pretraining. The third is further scaling, which raises the model’s parameter size to 400 million.

The LimiX team proposed CMNs as a new structured-data modeling paradigm that moves from “target prediction” toward “structure discovery.” Traditional PFNs center on the predictive distribution of a specified target; CMNs model the joint dependencies of all variables in the context data, focusing not only on what values variables take but also on how variables relate to one another and what internal structure they jointly form.

That shift resets the model’s learning objective. LimiX uses Contextual Conditional Masked Modeling, or CCMM, to organize cross-variable supervision under different observation patterns and to make variable-relationship inference an explicit pretraining target. The model no longer learns only information relevant to a specified outcome; through multiple conditional predictions, it jointly constrains the representation of dependencies across all variables.

The same principle runs through the network architecture. LimiX uses cell-level modeling, treating observations of variables in each sample as the basic representation unit. It preserves column identity, concrete values and missing states, and supports cross-variable and cross-sample information exchange. From data representation and training objectives to network computation, the design is organized around joint dependency modeling. Under this framework, classification, regression and missing-value imputation are unified as different queries to the same data model. Joint relationship modeling also provides a statistical basis for causal skeleton discovery when combined with causal assumptions. The team describes the goal as general data intelligence oriented toward inducing data mechanisms, rather than only a stronger predictor.

On training data, LimiX-2 upgrades a large-scale automated synthetic-data generation engine. Structured data is widely held in enterprise internal systems and is difficult to obtain publicly at the scale of internet text, so table foundation models such as TabPFN and Mitra also commonly use synthetic data for pretraining. LimiX-2 can automatically generate data with linear, nonlinear, multivariate interaction, periodic variation and noise perturbation distributions, exposing the model to more data structures and variable relationships before it enters real enterprise settings.

After Wenzhun Intelligent said in November last year that it had discovered and proposed that structured-data foundation models follow a scaling law, the new model continues along that route and increases parameters to 400 million. The increase is intended to validate the chosen technical route at a larger scale, improve generalization, and further exploit advantages in mining causal relationships and expressing them stably.

Beyond the prediction abilities reflected in the benchmarks, LimiX-2 also shows technical progress in causal discovery. Traditional causal discovery methods rely on expert priors or strict statistical assumptions and struggle with high-dimensional enterprise data. With the innovations above, LimiX-2 significantly leads traditional causal discovery methods on several public standard datasets, including Sachs, UF and Causal Chamber, according to QbitAI.

As of the 2026 World Artificial Intelligence Conference, the previous generation, LimiX, had completed generality validation in more than 3,800 structured-data scenarios and reached strategic cooperation with more than 60 customers. In earlier validations in real scenarios, LimiX changed the traditional machine-learning approach of training one dedicated model for one task into general prediction by a single foundation model across scenarios and datasets. Typical application directions include process parameter optimization, equipment failure prediction, electricity forecasting, energy consumption optimization and materials science.

Competition in the field is intensifying. Similar to the large language model track, structured-data foundation models are approaching a “GPT-3 moment,” according to QbitAI. Google, Amazon, SAP and multiple other well-known teams at home and abroad are accelerating model research and development. Chinese technical teams such as LimiX are continuing to refine their self-developed technical route and to explore a China model in which industrial scenarios drive model iteration.

Cui Peng, chief scientist of Wenzhun Intelligent, said: “The release of LimiX-2 and this Benchmark result are an important stage achievement for the team’s continued exploration in structured-data large models. We will still adhere to this technical route over the long term and continuously improve the model’s understanding of data patterns and variable causal relationships. In the future, we will continue to promote the model’s development in the direction of causal intelligence, helping artificial intelligence move faster from prediction to decision.”

The technical report is titled “LimiX-2: A Contextual Mechanism Network Towards General Structured-Data Intelligence.” It is available at https://arxiv.org/abs/2609.17488. The project’s GitHub page is https://github.com/limix-ldm-ai/LimiX, its Hugging Face page is https://huggingface.co/stable-ai/LimiX-2, and its ModelScope page is https://modelscope.cn/models/stable-ai/LimiX-2.