# TabICLv2: A better, faster, scalable, and open tabular foundation model

**Authors:** Jingang Qu, David Holzmüller, Gaël Varoquaux, Marine Le Morvan  
**arXiv:** [2602.11139](https://arxiv.org/abs/2602.11139)  
**Date:** 2026-02-11  
**Summary page:** https://www.summarizepaper.com/en/paper/2602.11139v1/

## Key points

- Tabular foundation models like TabPFNv2 and TabICL have risen to prominence, outperforming traditional gradient-boosted trees.
- The significance of in-context learning tailored for tabular datasets is highlighted by this shift.
- TabICLv2 is a cutting-edge foundation model for regression and classification with three key pillars of innovation:
- Novel synthetic data generation engine for high pretraining diversity
- Architectural enhancements including scalable softmax in attention mechanism for improved generalization capabilities
- Optimized pretraining protocols shifting from AdamW to the Muon optimizer
- On benchmark tests such as TabArena and TALENT, TabICLv2 surpasses RealTabPFN-2.5 without tuning, demonstrating remarkable generalization abilities on million-scale datasets within memory constraints while processing faster.
- Extensive ablation studies quantify the impact of each enhancement introduced in TabICLv2.
- Authors have released inference code and model weights on GitHub with plans to share synthetic data engine and pretraining code in subsequent releases.

## Summary

In the realm of predictive modeling for tabular data, recent advancements have seen Tabular foundation models like TabPFNv2 and TabICL rise to prominence, outperforming traditional gradient-boosted trees. This shift underscores the significance of in-context learning specifically tailored for tabular datasets. Enter TabICLv2, a cutting-edge foundation model for regression and classification that stands on three key pillars of innovation. Firstly, TabICLv2 boasts a novel synthetic data generation engine meticulously crafted to ensure high pretraining diversity. This engine sets the stage for robust model training and performance optimization. Secondly, the model incorporates various architectural enhancements, including a revolutionary scalable softmax in attention mechanism that enhances generalization capabilities across larger datasets without requiring prohibitively long-sequence pretraining. These innovations collectively contribute to improved model efficiency and accuracy. Moreover, TabICLv2 adopts optimized pretraining protocols, with a notable shift from AdamW to the Muon optimizer. This strategic change further refines the model's training process, resulting in enhanced performance outcomes. On benchmark tests such as TabArena and TALENT, TabICLv2 showcases its prowess by surpassing the current state-of-the-art RealTabPFN-2.5 without any tuning required. Notably, even with moderate pretraining compute resources, TabICLv2 demonstrates remarkable generalization abilities on million-scale datasets within 50GB GPU memory constraints while also exhibiting faster processing speeds compared to RealTabPFN-2.5. To substantiate these claims and contributions further, extensive ablation studies have been conducted to quantify the impact of each enhancement introduced in TabICLv2. In a commitment to open research practices, the authors have released inference code and model weights on GitHub (https://github.com/soda-inria/tabicl), with plans to share synthetic data engine and pretraining code in subsequent releases. Authored by Jingang Qu, David Holzmüller, Gaël Varoquaux, and Marine Le Morvan, "TabICLv2: A better, faster, scalable, and open tabular foundation model" represents a significant leap forward in predictive modeling for tabular data analysis.