GaiaChess 4.2.3
Download
One binary per platform. Windows binaries end in .exe, Linux and macOS have no extension.
On x86-64 the engine selects its SIMD paths at startup from your CPU — AVX2 up to AVX-512+VNNI for the
neural network, PEXT or AVX2 for move generation — so there is nothing to choose.
| Binary | Platform | Runs on |
|---|---|---|
universal |
Windows / Linux x86-64 | Any CPU from ~2013 (Intel Haswell+, AMD Zen+) |
compat |
Windows / Linux x86-64 | Any x86-64 CPU (2003+) |
neon |
Linux ARM64 | RPi 4+, Graviton... |
android-neon |
Android ARM64 | Android chess GUIs (DroidFish...), Termux |
apple-silicon |
macOS | M1 and later |
Take universal. Take compat only if universal does not start (a CPU older
than ~2013): it is the same engine on portable code paths — slower search, and a
scalar network below AVX2 — but it plays the same chess.
gaiachess info prints which paths were selected on your machine. The environment
variables GAIA_SIMD=scalar|avx2|avx512|vnni512 and GAIA_PEXT=0|1 cap the
selection, for benchmarking or troubleshooting.
Running on Linux
The binary is both the engine and the board: started on its own it opens the
interface, and a chess GUI that speaks UCI to it gets the engine instead.
Because it carries the interface's sound, Linux binaries need libasound2
(alsa-lib on Arch) — present on any desktop, sometimes missing from minimal
server images and containers. Windows and macOS need nothing extra.
android-neon is the exception: engine only (no board, no sound), statically
linked against musl, so it needs nothing at all — point your Android chess
GUI at the file and it runs.
All x86-64 binaries are compiled with PGO (Profile-Guided Optimization); every
SIMD tier the binary can select is profiled, the AVX-512 ones under Intel SDE.
The neural network is embedded — no external file needed.
Play in a browser
The whole game is a web page too, with nothing to install:
https://jromang.itch.io/gaiachess
gaiachess-web.zip is that page, for anyone who would rather serve it from
their own static host. Two things it wants: an embed of 528x630 (three times
the artwork, so nothing is scaled), and cross-origin isolation —
COOP/COEP headers — without which a search can only be abandoned, never
cut short.