Wenzhun Zhineng and Tsinghua Release LimiX-2, a 400M Foundation Model for Structured Data

On September 16, Wenzhun Zhineng and Tsinghua University released LimiX-2, a new generation data foundation model whose parameter scale has been raised to 400M. According to QbitAI, the model, built for structured data, topped an international evaluation leaderboard.

LimiX-2 is positioned as a foundation model for structured data. Unlike the general-purpose models that have dominated the past two years of natural language and multimodal work, this class of model operates on tables, database records and metric fields, the data that enterprises and institutions have accumulated for the longest and in the largest volumes.

A parameter scale of 400M is the clearest technical figure in the release. In a field where general models are routinely described in hundreds of billions of parameters, that number is itself a signal: work on structured data rewards specialised design for a particular data shape rather than raw scale.

The second piece of information in the report is the leaderboard result. For a foundation model, a public benchmark placement is a passport into enterprise procurement conversations, and in a domain that plugs directly into business systems, benchmark evidence tends to carry more weight than marketing language.

The work is a joint effort between Wenzhun Zhineng and Tsinghua University. Academia supplies method exploration while the company handles engineering and productisation, a combination that has become common in China's foundation model work and one reason vertical models can reach deployment relatively quickly.

The appeal of structured data is the density of use cases. Credit risk, corporate performance analysis, manufacturing operations and government data governance all depend on predictive modelling over tabular data, and traditional pipelines usually require per-task feature engineering and training. If general capability holds, the cost that can be squeezed out is substantial.

Three things are worth watching next: whether LimiX-2's weights or API are opened up, whether the leaderboard result can be reproduced by third parties, and whether it reliably beats existing task-specific models on real enterprise data. Those three answers decide whether this is a leaderboard moment or an industrial one.