0.17.1 (2026-02-17)

Performance

• Removed unnecessary and costly calls to typing.get_type_hints.

0.17.0 (2026-01-13)

Bug fix

• Using spox.inline with disabled value propagation (from spox._future) no longer raises an exception.

New features

• Added support for version 23 and 24 of opset ai.onnx.
• Added support for the following data types via the ml_dtypes package:
  • bfloat16
  • float8_e4m3fn
  • float8_e4m3fnuz
  • float8_e5m2
  • float8_e5m2fnuz
  • uint4
  • int4

Other change

• Spox now requires at least Python 3.10.

0.16.0 (2025-06-20)

New feature

• The instantiation time of constant numerical tensors has been significantly reduced by encoding their payload as raw bytes.

0.15.0 (2025-06-06)

Bug fix

• All models (including those used for value propagation operations) are now consistently built using IR version 8 introduced in ONNX 1.16.0.

0.14.1 (2025-06-04)

Technical release to address a PyPI publishing issue.

0.14.0 (2025-06-03)

Breaking change

• The spox.opset.ai.onnx.v18.group_normalization constructor function
  and its re-exports in spox.opset.ai.onnx.v19,
  spox.opset.ai.onnx.v20 have been removed due to it being deprecated
  upstream.

Bug fix

• Adds missing shape inference logic for spox.opsets.ai.v19.loop and
  spox.opsets.ai.v21.loop.

New feature

• The ai.onnx opset version 22 (released with ONNX 1.18) has been
  added to Spox.
• Propagated values may now be garbage collected if their associated
  Var object goes out of scope.

Other changes

• spox.opsets.ai.v17.loop, spox.opsets.ai.v19.loop and
  spox.opsets.ai.v21.loop will only infer shapes for loop carried
  dependencies if their shapes are unchanged across iterations.

0.13.0 (2024-12-06)

Support change

• Support for Python 3.8 has been dropped.

Bug fix

• Value propagation of string tensors no longer raises an erroneous ValueError in some instances.

Deprecation

• Importing spox._future.operator_overloading now triggers a DeprecationWarning. The function will be removed in a future release. Consider using ndonnx as an alternative.

Other changes

• The adaption logic is not using inferred values as initializers anymore.

0.12.1 (2024-06-18)

Bug fix

• Unset optional inputs are no longer erroneously prefixed by spox.inline.

Other changes

• The node-naming algorithm now has constant rather than quadratic time complexity.

0.12.0 (2024-05-16)

New feature

• The spox.build function gained the optional drop_unused_inputs argument.

0.11.0 (2024-04-23)

New feature

• Add opsets ai.onnx version 21 and ai.onnx.ml version 5 (released with ONNX 1.16)