Python 3.15's JIT Compiler Is Back on Track — 11% Faster on ARM

2026-03-18T11:33:37.000Z·1 min read
CPython's JIT, which was slower than the interpreter in Python 3.13-3.14, now delivers 11-12% speedups on macOS ARM and 5-6% on Linux x86_64. After losing its main sponsor, community stewardship and breaking problems into small tasks saved the project.

After a rocky start where the CPython JIT was often slower than the interpreter in Python 3.13 and 3.14, the project has hit its performance goals over a year early. The 3.15 alpha JIT delivers 11-12% speedups on macOS AArch64 and 5-6% on x86_64 Linux compared to the standard interpreter.

The Comeback Story

Ken Jin, a volunteer contributor, described the journey as anything but a heroic tale. The original CPython JIT in 3.13/3.14 showed practically no speedups, and the Faster CPython team lost its main sponsor in 2025 — casting doubt on the JIT's future.

The turnaround came through community stewardship and a deliberate strategy of breaking complex problems into manageable parts. Instead of massive architectural overhauls, the team opened "mega-issues" that split optimization work into simple, actionable tasks like "try optimizing a single instruction in the JIT."

What Changed

Performance Range

The geometric mean numbers (11-12% ARM, 5-6% x86_64) mask a wide range: from a 20% slowdown to over 100% speedup depending on the workload. These are preliminary numbers from the alpha release.

Free-Threading

Free-threading support (no-GIL) is not yet implemented but is targeted for 3.15/3.16, which would make the JIT performance gains available to multi-threaded Python code.

Key Contributors

Savannah Ostrowski, Mark Shannon, Diego Russo, Brandt Bucher, Ken Jin, Hai Zhu, Zheaoli, Tomas Roun, Reiden Ong, Donghee Na, and more — many of whom are volunteers stepping up after the sponsor withdrawal.

Source: Ken Jin's Blog | Benchmarks

↗ Original source
← Previous: Get Shit Done (GSD): Meta-Prompting System for Reliable AI Code GenerationNext: OnPrem.LLM: Launch Autonomous AI Agents with Sandboxed Execution in 2 Lines of Code →
Comments0