Wenzhun Intelligent and Tsinghua release LimiX-2, a 400M structured-data model topping global benchmarks

The model's overall Elo scores reached 1935 on TabArena, 1432 on BCCO and 1506 on TALENT, all first on the leaderboards, ahead of comparable models from Google, SAP and Amazon, Leiphone reported. On TabArena, LimiX-2 also recorded the best Elo in regression, binary classification and multiclass classification. TabArena, BCCO and TALENT are mainstream structured-data benchmarks covering multiple task types; they are used to measure models' generalization and prediction performance across different datasets.

The upgrade followed three technical routes. The first is Contextual Mechanism Networks, or CMNs, a prediction-dependency system that moves modeling focus from a pre-specified target variable to relationships among variables. The second is an upgraded automated synthetic-data generation engine used during pre-training. The third is further scaling, which raised the model parameter count to 400 million.

The LimiX team proposed CMNs to open a new structured-data modeling paradigm from target prediction to structure discovery. Traditional prior-fitted networks center on the predictive distribution of a specified target; CMNs instead model the joint dependencies of all variables in context data, attending not only to what values variables take but to how variables relate and what internal structure they form together. The change redefines the model's learning objective.

LimiX uses contextual conditional masked modeling, or CCMM, to organize cross-variable supervision under different observation modes, making variable-relationship inference an explicit pre-training goal. The model no longer learns only information relevant to a specified outcome; multiple conditional predictions jointly constrain the representation of all-variable dependency structure. This principle runs through the network architecture. LimiX adopts cell-level modeling, using observations of variables in each sample as the basic representation unit, preserving column identity, specific values and missing status, and supporting cross-variable and cross-sample information interaction. From data representation and training objective 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. The joint relationship modeling also provides a statistical basis for causal skeleton discovery when combined with causal assumptions. CMNs point not only to a stronger predictor but to general data intelligence oriented toward structural discovery and data-mechanism induction.

For training data, LimiX-2 upgraded a large-scale automated synthetic-data generation engine. Structured data exist widely inside enterprise systems and are difficult to obtain publicly at internet-text scale, so table foundation models such as TabPFN and Mitra also commonly use synthetic data for pre-training. LimiX-2 can automatically generate data under different distributions, including linear, nonlinear, multivariate interaction, periodic variation and noise perturbation, exposing the model to more data structures and variable relationships before it enters real enterprise settings.

After Wenzhun Intelligent identified and proposed in November last year that structured-data foundation models follow scaling laws, the new model continued along that technical route and increased parameters to 400 million. The larger scale is expected to validate LimiX's chosen route at a larger scale, improve generalization, and further exploit advantages in mining and stably expressing causal relationships.

Beyond prediction abilities reflected in the benchmarks, LimiX-2 showed a technical breakthrough in causal discovery. Traditional causal discovery methods rely on expert priors or strict statistical assumptions and struggle with high-dimensional enterprise data. With the new technical routes, LimiX-2 significantly leads traditional causal discovery methods on several public standard causal-discovery datasets, including Sachs, UF and Causal Chamber.

As of the 2026 World Artificial Intelligence Conference, the previous-generation LimiX had completed validation on more than 800 public datasets across industries and scenarios and achieved deployment in more than 60 real scenarios. In earlier real-world validations, LimiX turned the traditional machine-learning approach of training one specialized model per task into a general prediction capability of one foundation model across scenarios and datasets. Typical application directions include process parameter optimization, equipment failure prediction, power forecasting, energy consumption optimization and materials science.

Like the LLM track, structured-data foundation models are approaching a GPT-3 moment, and international competition is intensifying. Google, Amazon, SAP and other well-known teams at home and abroad are accelerating model development, while Chinese technical teams such as LimiX continue to refine their own technical routes and explore a China model in which industrial scenarios drive model iteration.

Cui Peng, chief scientist at Wenzhun Intelligent, said: 'The release of LimiX-2 and these benchmark results are an important milestone for the team's continued exploration in structured-data large models. We will stick to this technical route for the long term and keep improving the model's understanding of data patterns and variable causal relationships. In the future, we will continue to advance the model in the direction of causal intelligence to help AI move faster from prediction to decision-making.'

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

Editor's Summary Wenzhun Intelligent and Tsinghua University released LimiX-2, a 400-million-parameter structured-data foundation model that topped TabArena, BCCO and TALENT and reported gains in causal discovery. The release underscores intensifying competition in structured-data foundation models and the company's push from prediction toward decision-making. The technical report and model are publicly available.