NewYour coding agent can read the release notes before it upgrades.Set up the MCP server →
PyPI, #2393 most downloaded on PyPI
Flax: A neural network library for JAX designed for flexibility
Latest stable
0.12.12
Last release 25 days ago
27 Oct 2026
Ships fairly regularly
a new release about every 5 weeks
Nearly every release is documented
notes for 60 of the last 60 stable releases
4 versions withdrawn
withdrawn after publishing
7 years old
69 releases · first in 2020
Drag across the ruler to compare, or tap a release to open it.
Fixed an overly aggressive flax.nn deprecation warning in v0.3.1 that fired whenever you imported flax.
Fixed an overly aggressive flax.nn deprecation warning in v0.3.1 that fired whenever you imported flax.
flax.nn deprecation message no longer appears if you import flax directly.
NOTE: You must now explicitly import flax.nn if you want to use the old
pre-Linen flax.nn.Module.
Many improvements to Linen, and the old flax.nn is officially reprecated!
Many improvements to Linen, and the old flax.nn is officially reprecated!
Notably, there's a clean API for extracting intermediates from modules
defined using @nn.compact, a more ergonomic API for using Batch Norm and Dropout in modules
defined using setup, support for MultiOptimizer with Linen, and multiple safety, performance
and error message improvements.
See the CHANGELOG for more details
Your coding agent can read these notes before it upgrades. Set up the MCP server →
Many improvements to Linen, and the old flax.nn is officially deprecated!
Notably, there's a clean API for extracting intermediates from modules
defined using @nn.compact, a more ergonomic API for using Batch Norm and Dropout in modules
defined using setup, support for MultiOptimizer with Linen, and multiple safety, performance
and error message improvements.
Possible breaking changes:
Module instances are now frozen after setup has been called.
Previously mutations after setup could be dropped silently. Now the stateless requirement
is enforced by raising a TypeError in __setattr__ after setup.flax.core.apply and Module.apply. Now it returns a tuple
containing the output and a frozen empty
collection when mutable is specified as an empty list.broadcast_dims is now an attribute to Dropout instead of a __call__
argument.use_running_average and deterministic no longer have a default. They
should be passed explicitlyScope.variable mutability check, before a variable could only be
initialized if the 'params' collection was mutable.Other Improvements:
lm1b language modeling exampleflax.nn classes and methods.__call__. See design notesow method to Module and capture_intermediates argument to Module.apply.
See howto for usage patterns.variable argument to Module.apply to be a FrozenDictModelParamTraversal
As a result MultiOptimizer can be used properly with linen modules.is_mutable() method to Variable and is_mutable_collection() to flax.linen.Module.axis_name arg to flax.linen.vmapflax.linen.scanSee changelog, overall linen API improvements as well as a few bug fixes.
See changelog, overall linen API improvements as well as a few bug fixes.
Linen is now out of Alpha (flax.nn is being deprecated)!
flax.core.apply and linen Module.apply will now only return the variables
collections that were specified as mutable.Various bug fixes and new features, including the Linen API, a new functional core and many Linen examples.
Various bug fixes and new features, including the Linen API, a new functional core and many Linen examples.
Nothing published for this version
Minor update to push some fixes to pypi.
Minor update to push some fixes to pypi.
Add ConvTranspose Module to nn.linear
lhs_dilation -> input_dilation,
rhs_dilation -> kernel_dilationAdd ConvTranspose Module to nn.linear
Rename the following optional arguments to nn.linear.Conv: lhs_dilation -> input_dilation , rhs_dilation -> kernel_dilation
Change default layer names from numbers '0', '1', etc. to include the Module class name, e.g. 'Dense_0', 'LayerNorm_1'.
You can’t perform that action at this time.
Nothing published for this version
Nothing published for this version
One column per quarter.