Inria SODA

weights

open(v1, v1.1 and v2 checkpoints on the Hub, ungated)

data

open(synthetic prior regenerated by the released generator)

code

open(pretraining and synthetic data generation for v1 and v2)

license

BSD-3-Clause(code and weights)

Code and checkpoints are both BSD-3-Clause, and nothing is withheld: the pretraining corpus is synthetic, and the generator that produces it ships with the training code. The authors note that the released pretraining code was ported from their private codebase rather than being the exact code that trained the checkpoints. The time-series forecasting module is derived from Prior Labs' TabPFN-TS and carries Apache-2.0 inside the same license file.

• https://arxiv.org/abs/2602.11139 recorded 2026-09-26

  TabICLv2 abstract: "foster open research by releasing code for inference, pretraining, and synthetic data generation".

• https://cdn.jsdelivr.net/gh/soda-inria/tabicl@main/LICENSE recorded 2026-09-26

  LICENSE opens "BSD 3-Clause License", "Copyright (c) 2025, Soda team @ Inria"; a second section puts src/tabicl/forecast, derived from TabPFN-TS, under Apache License 2.0.

• https://cdn.jsdelivr.net/gh/soda-inria/tabicl@main/README.md recorded 2026-09-26

  README: "Pre-training code (including synthetic data generation) is available for both TabICLv1 and TabICLv2." and "By default, tabicl.train generates synthetic prior datasets on the fly ... this is how the TabICLv2 checkpoints were trained"; it also says the code "has been vibe-migrated from the original (private) pre-training codebase".

• https://huggingface.co/api/models/jingang/TabICL?expand[]=downloads&expand[]=cardData&expand[]=gated&expand[]=lastModified&expand[]=siblings recorded 2026-09-26

  Hub record jingang/TabICL: cardData.license bsd-3-clause, gated false, files tabicl-classifier-v1, v1.1 and v2 and tabicl-regressor-v2 checkpoints.

Measured on monthly PyPI downloads of tabicl, which downloads its checkpoints on first use; the Hub reports no downloads for the checkpoint repositories.

• https://pypistats.org/api/packages/tabicl/recent recorded 2026-09-26

  last_month 72,839 downloads of tabicl

TabICL answers the same question TabPFN does - a prediction on an unseen table with no training step - and its authors report it ahead of a tuned, ensembled RealTabPFN-2.5 on two public benchmarks. Like TabPFN it is one model, with no search or ensembling over others.

• https://arxiv.org/abs/2602.11139 recorded 2026-09-26

  TabICLv2 abstract: "On the TabArena and TALENT benchmarks, TabICLv2 without any tuning surpasses the performance of the current state of the art, RealTabPFN-2.5 (hyperparameter-tuned, ensembled, and fine-tuned on real data)."

• https://cdn.jsdelivr.net/gh/soda-inria/tabicl@main/README.md recorded 2026-09-26

  README: "It does not require hyperparameter tuning and outperforms heavily tuned XGBoost, CatBoost, or LightGBM on TabArena on ~80% of datasets."