Python 3.14.0 is now available (blog.python.org)

Python 3.14.0 is out, bringing a package of performance, concurrency and tooling changes that matter for AI/ML workflows. Headline items include official support for free-threaded (GIL‑less) builds via PEP 779 (enabling true concurrent threads for CPU-bound work, though opt-in and ecosystem adoption will be required), an experimental JIT shipped in official macOS and Windows binaries, and a new interpreter variant for newer compilers that can yield significantly better throughput when built from source. These runtime improvements, combined with improved error messages and faster UUID generation, directly target common ML pain points like model-serving latency, parallel data pipelines, and developer productivity. Several stdlib and tooling additions improve data handling, deployment and observability: PEP 784 adds a compression.zstd module for Zstandard (useful for model artifacts and datasets), PEP 734 exposes multiple interpreters in the stdlib (better isolation for multi-model hosts), PEP 649 defers annotation evaluation (cleaner typing/runtime separation), and PEP 768 introduces a zero‑overhead external debugger interface. Security and distribution also evolve: Sigstore is recommended for release verification (replacing PGP), official Android binaries are available, and the Windows install manager is being updated. Expect some incompatible changes and deprecations—test your ML stack before upgrading.