---
id: tabicl
version: "2.1.1"
license: BSD 3-Clause License Copyright (c) 2025, Soda team @ Inria Redistribution and use in source and binary forms, with or without modification, are permitted provided that the following conditions are… (full text in the JSON record)
license_treatment: permissive
maintenance: active
---

# tabicl — TabICL: A state-of-the-art tabular foundation model

License: permissive · Maintenance: active · Downloads: 142.9K/mo

## What it is and what it does

TabICL is a pre-trained transformer model designed to handle tabular classification and regression tasks without hyperparameter tuning. It works by performing in-context learning—learning from training data within a single forward pass—and can be used like a scikit-learn estimator with fit() and predict() methods. The model comes with pre-trained checkpoints and downloads them automatically on first use.

The package supports datasets ranging from hundreds to hundreds of thousands of samples with up to 2,000 features. It includes optional fine-tuning for single-dataset adaptation, KV caching to speed up repeated inference on the same training data, and CPU/disk offloading for larger datasets. Core dependencies are torch, numpy, scipy, scikit-learn, einops, huggingface-hub, psutil, and tqdm.

Use it for:

- Quick baseline predictions on tabular datasets without spending time tuning hyperparameters
- Classification or regression on datasets with hundreds to tens of thousands of samples where GPU inference is available
- Fine-tuning a pre-trained model on a single important dataset using FinetunedTabICLClassifier or FinetunedTabICLRegressor
- Repeated inference on the same training data with KV caching enabled to reduce latency
- Handling datasets with many features (up to 2,000) where traditional tabular models may struggle

## Worth the install?

AI-flagged interpretation of the facts above — verify before relying.

TabICL is a scikit-learn-compatible tabular foundation model that performs classification and regression on structured data via in-context learning, combining fit and predict in a single forward pass through a pre-trained transformer.

Yes, if you work with tabular data and want a zero-tuning baseline or a scikit-learn-compatible alternative to gradient boosting. The permissive license, low install friction, active maintenance, and no known vulnerabilities make it safe to try. Caveat: torch is a heavy dependency, and GPU access is recommended for the speed claims; CPU-only inference may be slower than claimed.

## Install

pip install tabicl
uv add tabicl
poetry add tabicl

## Installing tabicl

Before you install: Low friction install with a pure-Python wheel. Requires torch, numpy, scipy, scikit-learn, and several supporting libraries; torch installation on Intel Macs may require conda first. Actively maintained with recent releases.

License in practice: Dual-licensed under BSD 3-Clause (primary) and Apache 2.0 (for forecast module code derived from TabPFN-TS). Both are permissive; you may use, modify, and distribute freely with attribution and no warranty.

Quickstart:

pip install tabicl

from tabicl import TabICLClassifier
clf = TabICLClassifier()
clf.fit(X_train, y_train)
clf.predict(X_test)

torch installation on Intel Macs may require conda install pytorch -c pytorch before pip install tabicl

Verify before relying:

- Whether GPU acceleration is required for the claimed performance on 50,000 samples, or if CPU inference is practical
- Exact memory overhead of KV caching feature for typical dataset sizes
- Whether pre-trained checkpoint auto-download works reliably behind corporate proxies

## Package facts

- License: BSD 3-Clause License Copyright (c) 2025, Soda team @ Inria Redistribution and use in source and binary forms, with or without modification, are permitted provided that the following conditions are… (full text in the JSON record) (permissive)
- Python support: supports_current
- Install friction: low
- Maintenance: active
- Downloads: 142.9K/month (top 15,000 on PyPI)
- Known vulnerabilities: none known

## Tags

tabular classification without tuning, in-context learning tabular data, foundation model structured data, scikit-learn tabular predictor, transformer-based tabular regression, zero-shot tabular learning, tabular foundation model, tabular-learning, foundation-model, in-context-learning

[View on SkillFed](https://skillfed.io/packages/tabicl) · [View on PyPI](https://pypi.org/project/tabicl/)