What's new in scikit-learn 1.9: callbacks, sparse arrays, GPU support and the deprecations to fix

scikit-learn 1.9 came out in June 2026, and the bug-fix release 1.9.1 followed in September. It is the library behind most classic machine learning in Python: regression, classification, clustering and the pipelines around them. This release is less about new algorithms and more about the experience of using them: you can finally watch training progress, more models run on GPUs, and several long-standing quirks are fixed. It also starts the clock on a set of deprecations that become errors in 1.11.

pip install --upgrade scikit-learn

Callbacks: watch your models train

A long grid search used to be a black box: you started it and waited. scikit-learn 1.9 adds an experimental callback API, so estimators can report what they are doing during fit. It ships with two built-in callbacks, sklearn.callback.ProgressBar and sklearn.callback.ScoringMonitor, and supports nested progress, even when work runs in parallel.

Sparse arrays instead of sparse matrices

SciPy is moving from sparse matrices to sparse arrays, which behave more like NumPy arrays (for example, * means element-wise multiplication, not matrix multiplication). scikit-learn can now return either. The default is still "spmatrix", and the plan is to switch to "sparray" in a few releases, so it is worth trying now:

import sklearn
sklearn.set_config(sparse_interface="sparray")

# From here on, scikit-learn returns SciPy sparse arrays
# (for example csr_array) wherever it supports the new setting.
print(sklearn.get_config()["sparse_interface"])   # "sparray"

Faster and more accurate models

Also new: notebook displays of fitted estimators now show fitted attributes, and the ColumnTransformer view shows how output columns are assembled. narwhals is now a dependency and powers dataframe output, so set_output(transform="polars") and pandas output work more consistently.

More of scikit-learn runs on the GPU

Through the array API standard, scikit-learn can run some estimators directly on PyTorch or CuPy arrays, on a GPU, without converting to NumPy. 1.9 adds LogisticRegression and PoissonRegressor (with the lbfgs solver), Nystroem, FeatureUnion when all its steps support it, and more metrics such as average precision.

Deprecations to fix before 1.11

These raise warnings in 1.9 and will stop working in 1.11. Each has a simple fix:

# Deprecated: SVC(probability=True) is not thread-safe
from sklearn.svm import SVC
from sklearn.calibration import CalibratedClassifierCV

clf = CalibratedClassifierCV(SVC(), ensemble=False)
clf.fit(X_train, y_train)
proba = clf.predict_proba(X_test)

from sklearn.metrics import log_loss

# Deprecated: log_loss(y_test, proba) via y_pred
score = log_loss(y_test, y_proba=proba)

from sklearn.linear_model import LogisticRegressionCV

# The default scoring changes from accuracy to "neg_log_loss" in 1.11.
# Set it explicitly to keep today's behaviour and silence the warning.
model = LogisticRegressionCV(scoring="accuracy")

from sklearn.preprocessing import TargetEncoder
from sklearn.model_selection import KFold

# shuffle and random_state are deprecated: pass a CV splitter instead
enc = TargetEncoder(cv=KFold(n_splits=5, shuffle=True, random_state=0))

Upgrading your Python stack this autumn? Read the pandas 3.0 upgrade guide, what's new in Python 3.15 and set up clean environments with uv.

Sources: the scikit-learn blog "release 1.9" (12 June 2026) and the scikit-learn 1.9 release notes, including 1.9.1 (September 2026).

Alessandro Ecclesie AgazziWritten by Alessandro Ecclesie Agazzi, analytics engineer in London. Updated 17 October 2026. About me