NARC Genesys 5.1

NARC Genesys 5.1 is a small update, released two days after 5.0. During long time control testing after the 5.0 release, I found that the engine misjudged some positions and lost games because of it, mostly in longer games. 5.1 aims to fix this with a more accurate evaluation and better move ordering. The Elo figures below were measured at a short time control (8+0.08); results at longer time controls are still being collected.

Estimated strength: about 3608 CCRL Blitz (gauntlet against nine CCRL-rated engines at 8+0.08, 1,980 games). Against NARC Genesys 5.0: +29.9 ± 13.8 Elo (SPRT, 756 games at 8+0.08).

What's new in 5.1

• A wider network: 1024 neurons per perspective instead of 256, trained from scratch on the same 249M positions as 5.0, with the same labels from my own engines. It evaluates positions more accurately (validation loss 0.014254 against 5.0's 0.014808) but runs 25–40% slower; CPUs with a large cache lose the least speed.
• ThreatHist: the history that orders quiet moves is kept separately depending on whether the moving piece's square and its destination square are attacked, so better moves are tried first (+13.2 ± 12.7 Elo on its own, 892 games). setoption name ThreatHist value false gives the 5.0 search back.
• Networks of any hidden width (256, 512, 768 or 1024) load with EvalFile, so 5.0's network still works.

Which executable?

NARC-Genesys-5.1-windows.zip contains five Windows x86-64 builds. Each is statically linked with the network embedded, so no other files are needed.

File	CPU
NARC-Genesys-5.1-bmi2.exe	AVX2 + fast BMI2: Intel Haswell (2013) or newer, AMD Zen 3 or newer. Best choice for most computers.
NARC-Genesys-5.1-avx512.exe	AVX-512: Intel Ice Lake / Rocket Lake and newer, AMD Zen 4 and newer
NARC-Genesys-5.1-avx2.exe	AVX2 without fast BMI2: AMD Zen 1 / Zen 2 (Ryzen 1000–3000)
NARC-Genesys-5.1-sse42.exe	SSE4.2 + POPCNT: Intel Nehalem (2008) or newer
NARC-Genesys-5.1-generic.exe	any x86-64 CPU (slowest)

If unsure, try bmi2 first; if it closes at once or the GUI reports an error, go down the list. NARC Genesys has no GUI of its own: add it to a UCI GUI such as Arena, Cute Chess, Banksia or En Croissant. The .bmp next to each executable is its Arena logo.

UCI options

Option	Default
Hash	16	transposition table size in MB, transposition table size in MB; for longer time controls set 128–512 MB (more if you have the RAM)
Threads	1	more threads use Lazy SMP
Move Overhead	10	ms kept back per move for GUI or network delay
EvalFile	<embedded>	path of another .nnue network in the same format

Verification

NARC-Genesys-5.1-bmi2.exe bench        # signature: 1146093 nodes at depth 12
NARC-Genesys-5.1-bmi2.exe perftsuite   # move generation self-test on 7 reference positions

SHA-256 checksums of the zip and each executable are in NARC-Genesys-5.1.sha256.

License: GNU GPL v3 or later.

NARC Genesys 5.0

UCI chess engine in C++20 by Madushan Dissanayake, Sri Lanka. NARC Genesys is the fifth generation of the NARC engines and a complete rewrite of NARC Next 4.1, with an NNUE trained on about 249M positions labelled by the author's own engines.

Estimated strength: about 3600 CCRL Blitz (gauntlet against nine CCRL-rated engines at 8+0.08, 1,736 games).

Which executable?

NARC-Genesys-5.0-windows.zip contains five Windows x86-64 builds. Each is statically linked with the network embedded, so no other files are needed.

File	CPU
NARC-Genesys-5.0-bmi2.exe	AVX2 + fast BMI2: Intel Haswell (2013) or newer, AMD Zen 3 or newer. Best choice for most computers.
NARC-Genesys-5.0-avx512.exe	AVX-512: Intel Ice Lake / Rocket Lake and newer, AMD Zen 4 and newer
NARC-Genesys-5.0-avx2.exe	AVX2 without fast BMI2: AMD Zen 1 / Zen 2 (Ryzen 1000–3000)
NARC-Genesys-5.0-sse42.exe	SSE4.2 + POPCNT: Intel Nehalem (2008) or newer
NARC-Genesys-5.0-generic.exe	any x86-64 CPU (slowest)

If unsure, try bmi2 first; if it closes at once or the GUI reports an error, go down the list. NARC Genesys has no GUI of its own: add it to a UCI GUI such as Arena, Cute Chess, Banksia or En Croissant. The .bmp next to each executable is its Arena logo.

UCI options

Option	Default
Hash	16	transposition table size in MB
Threads	1	more threads use Lazy SMP
Move Overhead	10	ms kept back per move for GUI or network delay
EvalFile	<embedded>	path of another .nnue network in the same format

Verification

NARC-Genesys-5.0-bmi2.exe bench        # signature: 1405722 nodes at depth 12
NARC-Genesys-5.0-bmi2.exe perftsuite   # move generation self-test on 7 reference positions

SHA-256 checksums of the zip and each executable are in NARC-Genesys-5.0.sha256.

License: GNU GPL v3 or later.

NARC Next 4.1

Original C++20 UCI chess engine. Windows x86-64, statically linked, no DLLs
and no external network file required — the NNUE is embedded in the binary.

Which build to use

File	Requires	Notes
NARC-Next-4.1-x86-64-v3.exe	AVX2 / BMI2 (Haswell 2013+)	Recommended. ~25% faster.
NARC-Next-4.1-x86-64-v2.exe	SSE4.2 / POPCNT	Compatibility build for older CPUs.

Both are functionally identical and produce identical node counts at a fixed
depth; they differ only in speed. Neither is built with -march=native.

Protocol and options

UCI. Options: Hash (default 128 MB), Threads (default 1),
Move Overhead (default 30 ms), EvalFile (default <embedded>),
Use NNUE (default true), Use Policy (default false).

The engine does not learn between games and writes no persistent state.
There is no opening book in the binary.

Verification

printf 'uci\nisready\nposition startpos\ngo depth 13\n' | NARC-Next-4.1-x86-64-v3.exe

Expected at depth 13 from the start position: score cp 37, nodes 168609,
bestmove e2e4.

Note: the search runs on a background thread. Sending quit immediately after
go aborts it before completion and returns the first legal move; allow the
search to finish, as any GUI does.

Provenance

No Stockfish source code is copied or linked. The NNUE network is trained by
knowledge distillation on Stockfish 18 evaluation labels; the optional root
policy network (disabled by default) on Stockfish 8 move choices. Full
disclosure in the README.

License: GNU GPL v3 or later.