Version 1.6.0

We are happy to announce the AutoGluon 1.6.0 release!

AutoGluon 1.6.0 introduces new features and improvements to both the tabular and time series modules. This release adds several new tabular foundation models (TabPFN-3, TabDPT-Turbo, TabPFN-2.6, TabICLv2, and Nori) and the Toto-2 pretrained forecasting model, along with new presets that deliver strong accuracy and efficiency gains. It also includes new forecasting metrics, updated model memory estimation, refreshed dependency support (including transformers v5 and the latest PyTorch), and numerous bug fixes and quality-of-life improvements.

This release contains 184 commits from 19 contributors! See the full commit change-log here: https://github.com/autogluon/autogluon/compare/1.5.0…1.6.0

Join the community: Get the latest updates: Twitter

This release supports Python versions 3.10, 3.11, 3.12, and 3.13. Loading models trained on older versions of AutoGluon is not supported. Please re-train models using AutoGluon 1.6.0.

Spotlight

Tabular

  • New extreme preset, the recommended preset when a GPU is available: a 67% win-rate vs AutoGluon 1.5 extreme with 27x faster training and 4x faster inference, and an 84% win-rate vs best (4h) with 65x faster training. Led by the new Nori, TabICLv2, and TabDPT-Turbo foundation models alongside a TabPrep LightGBM, and every model in it is free for commercial use. @Innixma (#5812)

  • New noncommercial preset, setting a new state-of-the-art on TabArena: extreme plus TabPFN-3, outperforming TabFM in a quarter of the training time, with 20x faster inference and 10x fewer foundation model parameters. Free for research and internal experimentation; commercial use of TabPFN-3 requires a license from Prior Labs. @Innixma (#5812)

TabArena-Lite, 51/51 tasks, against the published leaderboard:

method

elo

rank

improvability (%)

train s

infer s

AutoGluon 1.6 noncommercial

1821

6.2

5.8

159

7.0

TabFM (default)

1791

7.0

5.1

613

138.4

AutoGluon 1.6 extreme

1745

8.5

6.8

137

6.2

AutoGluon 1.5 extreme (4h)

1638

12.9

7.9

3758

27.4

TabPFN-3 (default)

1632

13.2

9.1

26

2.4

AutoGluon 1.4 best (4h)

1516

20.0

12.1

8869

33.8

RealMLP (tuned + ensemble)

1480

22.5

13.0

31982

33.8

  • New validation_structure fit argument for non-IID data: declare a grouped or temporal structure and every internal validation split respects it across all models, keeping each group’s rows on one side of the split and validating strictly forward in time, with optional forward-chaining for temporal data. @Innixma (#5756, #5796)

  • New TabPrep feature engineering: the paper’s feature generators (group-by aggregates, random-subset compression, arithmetic and categorical interactions, out-of-fold target encoding) are now built in, and models can declare their own preprocessing pipelines via ag.model_specific_feature_generator_kwargs, applied per model rather than globally. This is what powers the TabPrep LightGBM in the new presets. @atschalz @Innixma @LennartPurucker (#5618, #5588, #5665, #5278)

General

Dependencies

Fixes and Improvements

Tabular

New Models

New Features

  • Remove fasttext model. @Innixma (#5582)

  • Remove old TabPFNv2 code. @Innixma (#5524)

  • Add core_kwargs and aux_kwargs fit params. @Innixma (#5646)

  • Express the auto-selected validation method as size curves. @Innixma (#5789)

  • Report constraint misses as skips, and add min_features / min_cells / max_cells. @Innixma (#5798)

  • Remove the feature cap on TabPFN-2.6 and TabPFN-3. @Innixma (#5797)

  • Lift Nori and TabDPT-Turbo feature caps, clamp Nori’s GPU estimate. @Innixma (#5801)

  • Apply TabICL’s pinned checkpoint, add a batch tier, lower the row cap. @Innixma (#5802)

  • Add support for quantile problem type in TabPFNv2. @colesussmeier (#5571)

  • Support loading Mitra from a local checkpoint path. @shchur (#5765)

  • Reduce feature importance mem usage by 25x. @Innixma (#5645)

  • GPU memory estimation plumbing + calibrated CPU/GPU memory estimates for foundation models. @Innixma (#5757)

  • Simplify the memory-estimate API: derive capability tags, default the instance wrappers. @Innixma (#5768)

  • Account for GPU memory when budgeting parallel bagging folds. @Innixma (#5791)

  • [experimental] Prototype GPU support for fit_strategy='parallel' (behind AG_PARALLEL_GPU). @Innixma (#5702)

  • Support xgboost v3.3: shared fit/predict category space for enable_categorical. @Innixma (#5784)

  • TabDPT: default to sequential fold fitting. @Innixma (#5785)

  • Minor changes for expert/experimental usage. @LennartPurucker (#5629)

API Changes and Deprecations

  • Make params immutable after construction (deprecation, raises in 1.7). @Innixma (#5776)

  • AuxiliaryParams typed schema; make params_aux immutable after construction (deprecation, raises in 1.7). @Innixma (#5775)

  • Declare supported problem types via _supported_problem_types class attribute. @Innixma (#5771)

  • Declare auxiliary-param overrides via _default_auxiliary_params_extra class attribute. @Innixma (#5772)

  • Declare ensemble-args overrides and fit-resource defaults as class attributes. @Innixma (#5777)

  • Validate _default_auxiliary_params_extra keys in verify_model. @Innixma (#5773)

Fixes and Improvements

TimeSeries

Highlights

  • Add Toto-2 pretrained forecasting model. @shchur (#5740, #5760)

  • New forecasting metrics: MAEB and WAPEB (bias-penalized), BIAS (to check for forecast bias), and MQL (mean quantile loss). @shchur (#5782, #5779)

  • New TimeSeriesPredictor.export_model method to export a trained model to a standalone checkpoint that loads without AutoGluon. @shchur (#5767)

  • New presets, achieving a 65% win rate vs. AutoGluon v1.5. @shchur (#5786, #5811)

New Features

  • [experimental] Add TimeSeriesPredictor.update() for ensemble re-selection on fresh data. @shchur (#5803)

API Changes and Deprecations

  • Prediction caching in TimeSeriesPredictor is now disabled by default and deprecated; it will be removed entirely in a future release. @shchur (#5761)

Fixes and Improvements

Multimodal

Fixes and Improvements

Documentation and CI

Contributors

Full Contributor List (ordered by # of commits):

@Innixma @shchur @celestinoxp @LennartPurucker @Atharva-Rajan-Kale @prateekdesai04 @jalencato @colesussmeier @atschalz @zhaozhiqiang200207 @petervanwylen @minkyu-choi07 @jwu7778 @felixkosmalla @eyh002 @bluehyena @biefan @MaximAliev @Mateusz-Switala

New Contributors

@zhaozhiqiang200207 @jwu7778 @felixkosmalla @MaximAliev @biefan @bluehyena @Mateusz-Switala @jalencato @Atharva-Rajan-Kale @petervanwylen @minkyu-choi07 @eyh002

Acknowledgements

AutoGluon’s continued development is made possible by the support of Amazon and Prior Labs. The project was created at Amazon, which has supported it since the very first release. With this release, Prior Labs joins in supporting AutoGluon’s core development, and the maintainer team now spans both organizations. We are grateful to both companies for investing in open source machine learning and enabling AutoGluon to keep pushing the state of the art.