v6.0.0

This release is a strength improvement of approximately 300 elo compared to version 5.1. This comes mostly from the greatly increased size of the neural network, with a hidden layer size of 384x2, 12x larger than the previous 32x2.

The release binaries are categorized by microarchitecture levels. The general rule of thumb is:

• x86-64-v3 will work on modern (2015+) CPUs (avx2, bmi1, lzcnt, etc.)
• x86-64-v2 will work on older (2009+) CPUs (popcnt, sse4, etc.)
• x86-64 will work on any x86_64 CPU.

v5.1.0

This release is a strength improvement of approximately 35 elo compared to version 5.0. This comes partially from improvements in search and partly from an improved neural network. This release will be the last one to feature the very small 32x2 hidden layer which has been used since version 3.0.

The release binaries are categorized by microarchitecture levels. The general rule of thumb is:

• x86-64-v3 will work on modern (2015+) CPUs (avx2, bmi1, lzcnt, etc.)
• x86-64-v2 will work on older (2009+) CPUs (popcnt, sse4, etc.)
• x86-64 will work on any x86_64 CPU.

v5.0.0

This release is a strength improvement of approximately 110 elo compared to version 4.0. This comes entirely from improvements in search; the neural network is the same as the one in version 4.0.

The release binaries are categorized by microarchitecture levels. The general rule of thumb is:

• x86-64-v3 will work on modern (2015+) CPUs (avx2, bmi1, lzcnt, etc.)
• x86-64-v2 will work on older (2009+) CPUs (popcnt, sse4, etc.)
• x86-64 will work on any x86_64 CPU.

v4.0.0

This release features a bucketed network architecture, ABDADA-based multithreading, and some search improvements. This release is approximately 100 elo stronger than version 3.0 in standard chess, and 160 elo stronger in FRC.

The release binaries are categorized by microarchitecture levels. The general rule of thumb is:

• x86-64-v3 will work on modern (2015+) CPUs (avx2, bmi1, lzcnt, etc.)
• x86-64-v2 will work on older (2009+) CPUs (popcnt, sse4, etc.)
• x86-64 will work on any x86_64 CPU.

v3.0.1

This release fixes a crash when the GUI sends a castling move in Chess 960 mode with the king on the E file and the rook on either the C or G files. This release is otherwise identical to version 3.0.0.

The release binaries are categorized by microarchitecture levels. The general rule of thumb is:

• x86-64-v3 will work on modern (2015+) CPUs (avx2, bmi1, lzcnt, etc.)
• x86-64-v2 will work on older (2009+) CPUs (popcnt, sse4, etc.)
• x86-64 will work on any x86_64 CPU.

v3.0.0

This release increases the size of the neural network (hidden layer size 32x2, up from 16x2) and further improves pruning during search. It also adds support for chess 960 (Fischer random chess). This release is approximately 220 elo stronger than version 2.1.0.

The release binaries are categorized by microarchitecture levels. The general rule of thumb is:

• x86-64-v3 will work on modern (2015+) CPUs (avx2, bmi1, lzcnt, etc.)
• x86-64-v2 will work on older (2009+) CPUs (popcnt, sse4, etc.)
• x86-64 will work on any x86_64 CPU.

v2.1.0

This release fixes a bug where Frozenight would repeat moves in winning positions. It also updates Frozenight's search algorithm to Principal Variation Search, which should help with Frozenight's relatively poor performance at longer time controls. Overall, this release is estimated to be about 50 elo stronger than version 2.0.0 at long time controls.

The release binaries are categorized by microarchitecture levels. The general rule of thumb is:

• x86-64-v3 will work on modern (2015+) CPUs (avx2, bmi1, lzcnt, etc.)
• x86-64-v2 will work on older (2009+) CPUs (popcnt, sse4, etc.)
• x86-64 will work on any x86_64 CPU.

v2.0.0

Most of the gains in this release come from a better, less buggy neural network. There are also some performance improvements and minor improvements to search. This release is approximately 250 elo stronger than version 1.0 at short time control (8s+0.08).

The release binaries are categorized by microarchitecture levels. The general rule of thumb is:

x86-64-v3 will work on modern (2015+) CPUs (avx2, bmi1, lzcnt, etc.)
x86-64-v2 will work on older (2009+) CPUs (popcnt, sse4, etc.)
x86-64 will work on any x86_64 CPU.

v1.0.1

This release fixes a bug where the move sent with bestmove uses chess960 castling notation instead of the standard UCI castling notation. This release is otherwise identical to version 1.0.0.

The release binaries are categorized by microarchitecture levels. The general rule of thumb is:

• x86-64-v3 will work on modern (2015+) CPUs (avx2, bmi1, lzcnt, etc.)
• x86-64-v2 will work on older (2009+) CPUs (popcnt, sse4, etc.)
• x86-64 will work on any x86_64 CPU.

v1.0.0

I think I've gotten pretty far in only two weeks of development and with such a small NNUE architecture.

The release binaries are categorized by microarchitecture levels. The general rule of thumb is:

• x86_64-v3 will work on modern (2015+) CPUs (avx2, bmi1, lzcnt, etc.)
• x86_64-v2 will work on older (2009+) CPUs (popcnt, sse4, etc.)
• x86_64 will work on any x86_64 CPU.