GitHub Repo
TensorFlow 2.22 Release Candidate Released, CUDA 13 Package Available and TensorBoard Dependency Changed
TensorFlow released 2.22.0-rc1 on November 17, with a dedicated CUDA 13 package also appearing on PyPI. This remains a pre-release, so engineering teams should first confirm package platform coverage and assess migration impacts from TensorBoard becoming a separate installation.
TensorFlow released 2.22.0-rc1 on November 17, advancing a new round of runtime and dependency changes to release-candidate status. The standard package was uploaded to PyPI starting on the 16th, and the dedicated CUDA 13 package became available on the 17th. Both are marked as pre-releases and should not yet be treated as stable releases. Official release notes, PyPI release history
The most direct migration change is that TensorBoard is no longer installed by default with TensorFlow. Applications using tf.summary.* or tf.keras.callbacks.TensorBoard must install tensorboard separately or they will encounter an ImportError. For teams using slim containers or rebuilding training environments, this means checking dependency manifests, not just verifying that models can be imported. Dependency change
For GPU deployments, the project lists initial CUDA 13 support through tensorflow-cuda13. At the time of checking, that package's rc1 files were limited to Linux x86-64 and covered CPython 3.11 through 3.14. The fact that the standard TensorFlow package provides Windows and macOS files does not mean CUDA 13 has the same platform coverage. This package split lets teams begin validating the new toolchain, but driver and workload compatibility still need to be tested. CUDA 13 package files
The rc1 notes also explicitly list support for standard and free-threaded Python 3.14 builds, and an upgrade of oneDNN to 3.14.1. CheckpointReader now releases the GIL while reading to address a multithreaded deadlock issue. These changes affect runtime environments and concurrency behavior; the release information does not provide benchmarks sufficient to estimate overall training speedups. Technical changes
One distinction matters: CUDA 13 and the TensorBoard dependency change were already mentioned in the earlier rc0 notes. The news here is the new release candidate and the availability of downloadable packages. Engineering teams should next monitor the stable release schedule, free-threaded package coverage, and regression results for existing training and monitoring workflows. Release-candidate history