Julia 1.13 Slashes Precompile Times, Overhauls REPL, and Speeds Up Garbage Collection

The latest release of the Julia language leans hard into responsiveness. Package precompilation runs about 30% faster than 1.12 and 10-20% faster than the 1.10 LTS, while cold startup improves roughly 20%. The team is now tracking Time-To-First-X systematically against 39 real-world community workflows, with dedicated CI jobs and public dashboards at perf.julialang.org/ttfx that went live September 7, 2026 — a shift from ad hoc benchmarking to continuous performance regression monitoring.

The interactive REPL gets a substantial upgrade: built-in syntax highlighting (no OhMyREPL.jl required), an fzf-style fuzzy history search that shows the originating REPL mode and supports multi-select, and bracketed paste support finally landing on Windows so pasting large blocks of code behaves correctly. A new public @FUNCTION macro resolves the innermost enclosing function even when anonymous, replacing reliance on the internal #self# variable.

Under the hood, two systems changes stand out. The default hash algorithm moves from MurmurHash3 (C) to RapidhashNano implemented in pure Julia — streaming, faster on long inputs (roughly 5x on a large string benchmark), and easier for custom types to opt into. The garbage collector now marks sysimage and package-image objects as permanently live and skips them during the mark phase, so full-collection cost scales with the heap a program actually allocates rather than with how much code is loaded — dropping a fresh-session full GC from tens of milliseconds to well under one. Notably, hash remains explicitly noncryptographic and the default seed has changed, so custom hash methods must accept the caller-supplied seed rather than defaulting it.