Repository navigation
Releases
v3.6.1
A forced move reports a searched score rather than a static evaluation. CHANGELOG.md has the measurements.
Which binary
The network is compiled into every binary, so each archive is self-contained.
| Archive | For |
|---|---|
windows-x86-64-avx2, linux-x86-64-avx2 |
Haswell/Zen 1 and newer. Pick this unless your CPU predates 2013 |
windows-x86-64, linux-x86-64 |
Any x86-64 with SSE4.2. About a quarter slower |
macos-arm64 |
Apple Silicon |
macos-x86-64-avx2 |
Intel Macs |
linux-aarch64-musl |
arm64 Linux, and Android via a terminal app. Static, so it needs nothing installed |
v3.6.0
Elementary mates. The engine converted 41 of 72 of them before the fifty-move rule; it now converts 69. CHANGELOG.md has the measurements.
This release also adds an arm64 Linux build.
Which binary
The network is compiled into every binary, so each archive is self-contained.
| Archive | For |
|---|---|
windows-x86-64-avx2, linux-x86-64-avx2 |
Haswell/Zen 1 and newer. Pick this unless your CPU predates 2013 |
windows-x86-64, linux-x86-64 |
Any x86-64 with SSE4.2. About a quarter slower |
macos-arm64 |
Apple Silicon |
macos-x86-64-avx2 |
Intel Macs |
linux-aarch64-musl |
arm64 Linux, and Android via a terminal app. Static, so it needs nothing installed |
v3.5.0
Search changes carried over from this engine's sibling project, and a score normalisation refitted for the current network. CHANGELOG.md has the measurements.
Which binary
The network is compiled into every binary, so each archive is self-contained.
| Archive | For |
|---|---|
windows-x86-64-avx2, linux-x86-64-avx2 |
Haswell/Zen 1 and newer. Pick this unless your CPU predates 2013 |
windows-x86-64, linux-x86-64 |
Any x86-64 with SSE4.2. About a quarter slower |
macos-arm64 |
Apple Silicon |
macos-x86-64-avx2 |
Intel Macs |
v3.4.0
A new network, and the inference rewritten around it. CHANGELOG.md has the measurements.
Which binary
The network is compiled into every binary, so each archive is self-contained.
| Archive | For |
|---|---|
windows-x86-64-avx2, linux-x86-64-avx2 |
Haswell/Zen 1 and newer. Pick this unless your CPU predates 2013 |
windows-x86-64, linux-x86-64 |
Any x86-64 with SSE4.2. About a quarter slower |
macos-arm64 |
Apple Silicon |
macos-x86-64-avx2 |
Intel Macs |
v3.3.0
Contempt is removed, and reported evaluations are normalised so that +1.00 means a 50% chance of winning. CHANGELOG.md has the measurements.
Which binary
The network is compiled into every binary, so each archive is self-contained.
| Archive | For |
|---|---|
windows-x86-64-avx2, linux-x86-64-avx2 |
Haswell/Zen 1 and newer. Pick this unless your CPU predates 2013 |
windows-x86-64, linux-x86-64 |
Any x86-64 with SSE4.2. About a quarter slower |
macos-arm64 |
Apple Silicon |
macos-x86-64-avx2 |
Intel Macs |
v3.2.0
Search fixes ported over from this engine's sibling project, plus the time-management redesign from v3.1.0. CHANGELOG.md has the measurements.
Which binary
The network is compiled into every binary, so each archive is self-contained.
| Archive | For |
|---|---|
windows-x86-64-avx2, linux-x86-64-avx2 |
Haswell/Zen 1 and newer. Pick this unless your CPU predates 2013 |
windows-x86-64, linux-x86-64 |
Any x86-64 with SSE4.2. About a quarter slower |
macos-arm64 |
Apple Silicon |
macos-x86-64-avx2 |
Intel Macs |
v3.0.0
The network now sees threats and pawn pairs, not just where the pieces stand. That is worth about 110 Elo over v2.1.0 at 10s+0.1s on one thread; CHANGELOG.md has the measurement.
Which binary
The network is compiled into every binary, so each archive is self-contained.
| Archive | For |
|---|---|
windows-x86-64-avx2, linux-x86-64-avx2 |
Haswell/Zen 1 and newer. Pick this unless your CPU predates 2013 |
windows-x86-64, linux-x86-64 |
Any x86-64 with SSE4.2. About a quarter slower |
macos-arm64 |
Apple Silicon |
The network is trained on data provided by the Leela Chess Zero project, made available under the Open Database License.