China’s LimiX-2 Tops Global Structured-Data Benchmarks as SAP, Google and Amazon Enter LDM Race

The team behind LimiX-2 is made up of Tsinghua University PhDs, according to QbitAI. Its chief scientist is Cui Peng, a professor in Tsinghua University’s Department of Computer Science and Technology, who has long worked on structured data and causal intelligence and is described as one of the proponents of the large data model, or LDM, concept. Zhunzhun Intelligent previously released LimiX, described by QbitAI as China’s first structured-data foundation model, which also held leading positions in international rankings for an extended period.

LDM targets structured data rather than text. According to the report, LDM aims to let foundation models learn variable relationships, conditional relationships and generative mechanisms directly from different structured datasets. Large language models have advanced rapidly in text, speech, images and video, but industrial production and scientific research require causality, accuracy and stability in decision-making, where structured data remains central. In manufacturing, dozens, hundreds or even thousands of variables such as temperature, pressure, rotational speed, raw-material ratios, equipment status and energy consumption change simultaneously, forming a continuously evolving high-dimensional data space. Many key patterns may not have been summarized by humans as rules, formulas or text knowledge.

QbitAI reported that converting such data into text before passing it to an LLM introduces an information-compression step. In production systems, fine-grained differences that appear negligible in semantic space may determine prediction accuracy, yield and risk judgment. LimiX-2 therefore models variable relationships as a core objective rather than treating them as intermediate information for prediction. The report says LimiX-2 uses Contextual Mechanism Networks, or CMNs, to move structured-data modeling from target-variable-centered prediction to relationship modeling of joint dependencies and data-generation mechanisms across all variables.

The training method includes Contextual Conditional Masking Modeling, or CCMM, and an upgraded automated synthetic-data engine. By repeatedly masking and predicting different variables, the model is trained to learn conditional dependencies and joint structures among variables rather than fitting a single fixed target. LimiX-2 also lowers its modeling granularity to the cell level, retaining column identity, specific values and missing states, and supporting interaction across variables and samples. Under this framework, classification, regression and missing-value imputation can be treated as different queries against the same data model, according to the report.

On TabArena, LimiX-2 ranked first on four indicators by Elo in classification tasks, reaching an Elo of 1917 and leading TabFM by 143 points. In regression tasks on TabArena, it also ranked first on four indicators by Elo, with an Elo of 2206. The report said LimiX-2 surpassed Google TabFM, TabPFN, TabICL and Mitra in these evaluations. The company also upgraded its large-scale automated synthetic-data engine to cover more distributions, interaction relationships and noise types, supporting the expanded learning objectives of CMNs.

Beyond prediction, LimiX-2 has moved into causal discovery. QbitAI reported that on public datasets including Sachs, UF and Causal Chamber, LimiX-2 outperformed several traditional causal-discovery methods. The report described classification, regression and cross-dataset generalization as daily enterprise tasks, and said such benchmarks test whether a model can make stable predictions, transfer across data and support decisions. This points to a longer-term path for LDM from prediction toward relationship modeling.

The LDM field has attracted major technology companies. According to QbitAI, Google released TabFM this year, formally adding tabular-data foundation models to its foundation-model portfolio. Amazon launched Mitra to explore how a pretrained model can transfer directly across different tables and tasks. SAP first built SAP-RPT-1, then acquired Prior Labs, which is behind TabPFN, and invested more than 1 billion euros to expand its structured-data AI research team. A few days before the report, TabPFN-3.5 Plus entered SAP AI Core for tasks including cash-flow forecasting, payment delays, supplier risk and customer churn.

The report contrasted LimiX-2’s approach with PFN-based routes represented by TabPFN. It said PFN models are designed to adapt quickly to a new table given context and a target and to predict Y accurately, which suits classification and regression. But when users need to know not only the value of Y but also relationships among variables and how changes in one variable affect others, the target-centered PFN route shows limits. For many PFN models, data is first compressed into a higher-level representation before target prediction, which helps rapid task adaptation but can fold away some variable-level fine-grained information.

QbitAI said the LimiX-2 team did not continue stacking capabilities along the traditional prediction paradigm but changed what the model learns. CMNs, CCMM and cell-level representation form a technical paradigm covering learning objectives, training mechanisms and data representation. The report compared the underlying shift to the move from BERT to GPT in large language models, saying the core change is not only parameter scale or benchmark scores but the redefinition of learning objectives and how information is organized.