Valkand 1.1 - new UCI engine (NNUE) for CCRL testing

[valkand66]

Hello,

I would like to submit Valkand 1.1 for testing in the CCRL lists.

Download (Windows, Linux, macOS; 64-bit; net built into the executable): https://github.com/valkand66/valkand-ch ... s/tag/v1.1

- Protocol: UCI. Options: Hash, Threads (1-256), Ponder, Move Overhead, Contempt (default 15 cp, repetitions only), SyzygyPath, OwnBook (small built-in book, off by default). Pondering (UCI Ponder) is supported; no MultiPV in this version.
- Binaries:
Windows x64 and Linux x64: avx2 (recommended) and popcnt versions.
Linux ARM64 (needs dotprod: Raspberry Pi 5, Graviton2+...; not Raspberry Pi 4).
macOS Apple Silicon (M1 to M6).
The neural net is embedded in the executable: no extra files are needed.
MD5 sums of every archive are in the README of the repository (https://github.com/valkand66/valkand-chess).
- Engine: written in C by Krassimir Kandjikov (Val Kand): bitboards, alpha-beta search, NNUE evaluation written from scratch
(768 features x 16 king zones -> 256 -> 16 -> 1, 8 output buckets, int8). Developed with the help of an AI assistant (Claude).
- Net: trained by me on ~830M positions labelled with Stockfish 16 (depth 14) evaluations and the Lichess evaluation database (CC0).
Positions come from my own self-play games and public Lichess games. No third-party net file is used.
- Syzygy probing via Fathom (MIT).
- Strength: I have not measured an absolute rating yet, so I prefer not to quote one; I would be glad to see what CCRL gives.
- License: Freeware, closed source (see the README in each archive). Contact: valkand66@gmail.com

Thanks!

[flok]

Hello,

I would like to submit Valkand 1.1 for testing in the CCRL lists.

Download (Windows, Linux, macOS; 64-bit; net built into the executable): https://github.com/valkand66/valkand-ch ... s/tag/v1.1

- Protocol: UCI. Options: Hash, Threads (1-256), Ponder, Move Overhead, Contempt (default 15 cp, repetitions only), SyzygyPath, OwnBook (small built-in book, off by default). Pondering (UCI Ponder) is supported; no MultiPV in this version.
- Binaries:
Windows x64 and Linux x64: avx2 (recommended) and popcnt versions.
Linux ARM64 (needs dotprod: Raspberry Pi 5, Graviton2+...; not Raspberry Pi 4).
macOS Apple Silicon (M1 to M6).
The neural net is embedded in the executable: no extra files are needed.
MD5 sums of every archive are in the README of the repository (https://github.com/valkand66/valkand-chess).
- Engine: written in C by Krassimir Kandjikov (Val Kand): bitboards, alpha-beta search, NNUE evaluation written from scratch
(768 features x 16 king zones -> 256 -> 16 -> 1, 8 output buckets, int8). Developed with the help of an AI assistant (Claude).
- Net: trained by me on ~830M positions labelled with Stockfish 16 (depth 14) evaluations and the Lichess evaluation database (CC0).
Positions come from my own self-play games and public Lichess games. No third-party net file is used.
- Syzygy probing via Fathom (MIT).
- Strength: I have not measured an absolute rating yet, so I prefer not to quote one; I would be glad to see what CCRL gives.
- License: Freeware, closed source (see the README in each archive). Contact: valkand66@gmail.com

Thanks!

Closed source, trained with stockfish, developed "with the help" of an ai.

[valkand66]

Thanks for looking, flok. All three points are correct, and I stated them openly in the post. Closed source is my choice for now. The net is trained on Stockfish evaluations used as labels (plus the Lichess evaluation database), a common technique; I don’t use any third-party net file. And yes, I wrote the engine with an AI assistant. If CCRL decides it doesn’t fit their lists for any of these reasons, I understand. Otherwise I’m happy to answer questions about the engine.

[Graham Banks]

A rough idea of rating would be useful.

[valkand66]

Thanks, Graham. I haven’t measured an absolute rating yet. I’m going to run a match against Stockfish and will post a rough figure here later today.

[valkand66]

Thanks, Graham. I ran a quick gauntlet (1 thread, 64 MB hash, 10+0.1, 200 games each) against three engines from your list: Lynx 1.11.0 (+147 Elo for Valkand), Odonata 1.1.0 (+129) and Nalwald 19 (+177). Taking their CCRL 40/15 ratings as reference, that points to roughly 3400–3450 on your scale, but the time control and hardware are very different from CCRL’s, so please take it as an upper-bound guess. Stockfish 19 is clearly stronger than Valkand in my tests. Happy to send a longer-TC match if useful.

[Graham Banks]

Thanks for giving a rough idea of your engine's strength.