Xiaomi unveils open-source tabldm model for structured data, leads regression benchmarks

On September 5, 2026, Xiaomi officially released its open-source general-purpose table data foundation model, Xiaomi-TabLDM. The model is designed to adapt to various table datasets with a single pre-training setup, eliminating the need for retraining or fine-tuning for each task. This approach addresses the inefficiencies of traditional table machine learning, which often requires repeated training and model integration. Xiaomi-TabLDM advances in three areas: pre-training with synthetic data generated by structural causal models (SCM), a novel model architecture featuring dual-stream feature grouping and sparse Mixture-of-Experts, and Test-Time Scaling for improved inference. It outperformed existing models in four benchmark tests, ranking first in regression on OpenML-CTR23 and second in TALENT, TabArena, and BCCO. In TabArena, it achieved the second-highest Elo score while reducing training time by 82% and prediction time by 68% compared to TabFM. In industrial applications, Xiaomi-TabLDM demonstrated significant improvements over traditional methods, achieving a 130% increase in prediction accuracy without fine-tuning and reducing testing time by 90%. It also improved prediction accuracy in parts weight and production composition forecasting, with error reductions of 31% and 54% respectively.