Version 1.5.0

We are happy to announce the AutoGluon 1.5.0 release!

AutoGluon 1.5.0 introduces new features and major improvements to both tabular and time series modules.

This release contains 131 commits from 17 contributors! See the full commit change-log here: https://github.com/autogluon/autogluon/compare/1.4.0…1.5.0

Join the community: Get the latest updates: Twitter

This release supports Python versions 3.10, 3.11, 3.12 and 3.13. Support for Python 3.13 is currently experimental, and some features might not be available when running Python 3.13 on Windows. Loading models trained on older versions of AutoGluon is not supported. Please re-train models using AutoGluon 1.5.0.

Spotlight

Chronos-2

AutoGluon v1.5 adds support for Chronos-2, our latest generation of foundation models for time series forecasting. Chronos-2 natively handles all types of dynamic covariates, and performs cross-learning from items in the batch. It produces multi-step quantile forecasts and is designed for strong out-of-the-box performance on new datasets.

Chronos-2 achieves state-of-the-art zero-shot accuracy among public models on major benchmarks such as fev-bench and GIFT-Eval, making it a strong default choice when little or no task-specific training data is available.

In AutoGluon, Chronos-2 can be used in zero-shot mode or fine-tuned on custom data. Both LoRA fine-tuning and full fine-tuning are supported. Chronos-2 integrates into the standard TimeSeriesPredictor workflow, making it easy to backtest, compare against classical and deep learning models, and combine with other models in ensembles.

from autogluon.timeseries import TimeSeriesPredictor

predictor = TimeSeriesPredictor(...)
predictor.fit(train_data, presets="chronos2")  # zero-shot mode

More details on zero-shot usage, fine-tuning and ensembling are available in the updated tutorial.

AutoGluon Tabular

AutoGluon 1.5 Extreme sets a new state-of-the-art on TabArena, with a 60 Elo improvement over AutoGluon 1.4 Extreme. On average, AutoGluon 1.5 Extreme trains in half the time, has 50% faster inference speed, a 70% win-rate, and 2.8% less relative error compared to AutoGluon 1.4 Extreme. Whereas 1.4 used a mixed portfolio that changed depending on data size, 1.5 uses a single fixed portfolio for all datasets.

Notable Improvements:

  1. Added TabDPT model, a tabular foundation model pre-trained exclusively on real data.

  2. Added TabPrep-LightGBM, a LightGBM model with custom preprocessing logic including target mean encoding and feature crossing.

  3. Added early stopping logic for the portfolio which stops training early for small datasets to mitigate overfitting and reduce training time.

AutoGluon 1.5 Extreme uses exclusively open and permissively licensed models, making it suitable for production and commercial use-cases.

To use AutoGluon 1.5 Extreme, you will need a GPU, ideally with at least 20 GB of VRAM to ensure stability. Performance gains are primarily on datasets with up to 100k training samples.

# pip install autogluon.tabular[tabarena]  # <-- Required for TabDPT, TabICL, TabPFN, and Mitra
from autogluon.tabular import TabularPredictor
predictor = TabularPredictor(...).fit(train_data, presets="extreme")  # GPU required
TabArena All (51 datasets)
Model Elo [⬆️] Improvability (%) [⬇️] Train Time (s/1K) [⬇️] Predict Time (s/1K) [⬇️]
AutoGluon 1.5 (extreme, 4h) 1736 3.498 289.07 4.031
AutoGluon 1.4 (extreme, 4h) 1675 6.381 582.21 6.116
AutoGluon 1.4 (best, 4h) 1536 9.308 1735.72 2.559
Pareto Frontier (Elo) Pareto Frontier (Improvability)
pareto_n_configs_elo pareto_n_configs_imp

New Model: RealTabPFN-2.5

Tech Report: TabPFN-2.5: Advancing the State of the Art in Tabular Foundation Models

AutoGluon 1.5 adds support for fitting the RealTabPFN-2.5 model, the current strongest individual model on TabArena. Unlike TabPFN-2 which has a permissive license, RealTabPFN-2.5 comes with a non-commercial license and requires the user to authenticate with HuggingFace and accept a terms of use agreement before being able to download the weights. The user will be automatically prompted to perform these steps during AutoGluon’s fit call if RealTabPFN-2.5 is specified, and the model will be skipped until the weights have been downloaded by the user. RealTabPFN-2.5 is not currently used in any AutoGluon preset, and must be manually specified.

All TabPFN user telemetry is disabled when used with AutoGluon.

To use RealTabPFN-2.5 (non-commercial use only):

# pip install autogluon.tabular[all,tabpfn]
from autogluon.tabular import TabularPredictor
predictor = TabularPredictor(...).fit(
    train_data,
    hyperparameters={"REALTABPFN-V2.5": [{}]},
)  # GPU required, non-commercial

To use RealTabPFN-2 (permissive license):

# pip install autogluon.tabular[all,tabpfn]
from autogluon.tabular import TabularPredictor
predictor = TabularPredictor(...).fit(train_data, hyperparameters={"REALTABPFN-V2": [{}]})  # GPU required

For users who previously were using "TABPFNV2", we strongly recommend switching to "REALTABPFN-V2" to avoid breaking changes in the latest TabPFN releases.

New Model: TabDPT

Paper: TabDPT: Scaling Tabular Foundation Models on Real Data

To use TabDPT (permissive license):

# pip install autogluon.tabular[all,tabdpt]
from autogluon.tabular import TabularPredictor
predictor = TabularPredictor(...).fit(train_data, hyperparameters={"TABDPT": [{}]})  # GPU recommended

New Model: TabPrep-LightGBM

TabPrep-LightGBM is an experimental model that uses a custom data preprocessing pipeline to enhance the performance of LightGBM. It represents a working snapshot of an in-progress research effort. Further details will be shared as part of an upcoming paper.

TabPrep-LightGBM achieves a new state-of-the-art for model performance on TabArena’s 15 largest datasets (10k - 100k training samples), exceeding RealTabPFN-2.5 by 100 Elo while fitting 3x faster using just 8 CPU cores. TabPrep-LightGBM is also incorporated into the AutoGluon 1.5 extreme preset.

TabArena Medium (10k - 100k samples, 15 datasets)

Model

Elo [⬆️]

Imp (%) [⬇️]

Train Time (s/1K) [⬇️]

Predict Time (s/1K) [⬇️]

AutoGluon 1.5 (extreme, 4h)

1965

1.876

191.18

2.207

AutoGluon 1.4 (extreme, 4h)

1813

3.016

289.53

3.187

AutoGluon 1.4 (best, 4h)

1794

3.122

432.35

4.085

TabPrep-LightGBM (tuned + ensembled)

1787

3.573

256.12

2.281

RealTabPFN-v2.5 (tuned + ensembled)

1680

5.818

735.58

11.736

RealMLP (tuned + ensembled)

1649

6.102

1719.82

1.675

ModernNCA (tuned + ensembled)

1636

6.189

2526.28

6.013

CatBoost (tuned + ensembled)

1616

6.011

777.59

0.25

LightGBM (tuned + ensembled)

1598

7.77

131.56

2.639

To use TabPrep-LightGBM, we recommend trying the presets it is used in: "extreme", "best_v150", "high_v150". Fitting TabPrep-LightGBM outside of the use of presets is currently complicated.

General

Dependencies

Fixes and Improvements

Tabular

AutoGluon-Tabular v1.5 introduces several improvements focused on accuracy, robustness, and usability. The release adds new foundation models, updates the feature preprocessing pipeline, and improves GPU stability and memory estimation. New model portfolios are provided for both CPU and GPU workloads.

Highlights

  • New models: RealTabPFN-2, RealTabPFN-2.5, TabDPT, TabPrep-LightGBM, and EBM are now available in AutoGluon-Tabular.

  • Updated preprocessing pipeline with more consistent feature handling across models.

  • Improved GPU stability and more reliable memory estimation during training.

  • New CPU and GPU portfolios tuned for better performance across a wide range of datasets: "extreme", "best_v150", "high_v150".

  • Stronger benchmark results: with the new presets, AutoGluon-Tabular v1.5 Extreme achieves a 70% win rate over AutoGluon v1.4 Extreme on the 51 TabArena datasets, with a 2.8% reduction in mean relative error.

New Features

Fixes and Improvements

TimeSeries

AutoGluon v1.5 introduces substantial improvements to the time series module, with clear gains in both accuracy and usability. Across our benchmarks, v1.5 achieves up to an 80% win rate compared to v1.4. The release adds new models, more flexible ensembling options, and numerous bug fixes and quality-of-life improvements.

Highlights

  • Chronos-2 is now available in AutoGluon, with support for zero-shot inference as well as full and LoRA fine-tuning (tutorial).

  • Customizable ensembling logic: Adds item-level ensembling, multi-layer stack ensembles, and other advanced forecast combination methods (documentation).

  • New presets leading to major gains in accuracy & efficiency. AG-TS v1.5 achieves up to 80% win rate over v1.4 on point and probabilistic forecasting tasks. With just a 10 minute time limit, v1.5 outperforms v1.4 running for 2 hours.

  • Usability improvements: Automatically determine an appropriate backtesting configuration by setting num_val_windows="auto" and refit_every_n_windows="auto". Easily access the validation predictions and perform rolling evaluation on custom data with new predictor methods backtest_predictions and backtest_targets.

New Features

API Changes and Deprecations

  • Remove outdated presets related to the original Chronos model: chronos, chronos_large, chronos_base, chronos_small, chronos_mini, chronos_tiny, chronos_ensemble. We recommend to use the new presets chronos2, chronos2_small and chronos2_ensemble instead.

Fixes and Improvements

Code Quality

Multimodal

Fixes and Improvements

Documentation and CI

Contributors

Full Contributor List (ordered by # of commits):

@shchur @canerturkmen @Innixma @prateekdesai04 @abdulfatir @LennartPurucker @celestinoxp @FANGAreNotGnu @xiyuanzh @nathanaelbosch @betatim @AdnaneKhan @paulbkoch @shou10152208 @ryuichi-ichinose @atschalz @colesussmeier

New Contributors