On September 5, 2026, Xiaomi formally released and open-sourced Xiaomi-TabLDM, a general-purpose tabular-data foundation model. With a single pretrained model and one default configuration it adapts to different tabular datasets without per-task retraining, tuning or ensembling, handling both classification and regression across finance, healthcare, manufacturing and logistics.
The 70M-parameter model is pretrained entirely on large-scale synthetic data generated from a structural causal model (SCM), covering varied data scales, variable types and functional dependencies. Its architecture introduces dual-stream feature grouping, a lightweight attention residual and sparse mixture-of-experts to model feature relations at multiple granularities, while Test-Time Scaling boosts prediction by adding inference compute without changing pretrained weights.
On four public benchmarks - TALENT, TabArena, BCCO and OpenML-CTR23 - Xiaomi-TabLDM reached the first tier. It ranked first on OpenML-CTR23 regression and on TALENT binary classification, and second on regression in the other three. On TabArena regression it posted the second-highest Elo score while cutting training time 82% and prediction time 68% versus the top-ranked TabFM.
In real industrial checks it lifted material-property prediction accuracy by 130% with no fine-tuning and cut无效 trial molding by about 90%; versus XGBoost it lowered failed-sample ratio by about 31% in part-weight prediction and average error by 54% in production-composition prediction, needing only about 30 new samples to cut error a further 62% after a condition change. Weights and code are open on GitHub and Hugging Face with a scikit-learn-compatible interface.
Image: Efficiency of AI-related computer chips. Source: Wikimedia Commons (CC BY-SA 4.0).
