New engine releases 2026 H2

[FlavusSnow]

FunComp 3.0

https://github.com/Moxnar/FunComp (release: https://github.com/Moxnar/FunComp/releases/tag/v3.0)

FunComp 3.0 is the first version with an NNUE evaluation. It is estimated at about 3640 on the CCRL Blitz scale (3642 ± 8 from a single-threaded gauntlet of 2,000 games at 10+0.1 against 20 rated engines, CCRL Blitz 3492–3793; an estimate on my hardware, not an official CCRL rating). That’s about +360 above version 2.0 measured the same way, and +484 ± 33 against 2.0 in direct self-play.

Relatively simple net: a single 1024-wide hidden layer with 7 king buckets and 8 output buckets. It was trained on ~1.44B positions from FunComp’s own self-play, labelled only by its own search, with no other engine’s games, evaluations or networks. The net is built into the executable, so there’s nothing extra to download.

Windows binaries for AVX2 CPUs and for any x86-64 CPU. Supports multiple threads, Syzygy tablebases and pondering. GPL-3.0.

What's next: Better training labels and more of them.

[Graham Banks]

Kociolek v2.3

https://github.com/Koci0l/NNUE-Chess-en ... s/tag/v2.3
Did 2 datagens. First 2.0B pos then on the new net 2.5B.
Elo: 54.29 +/- 21.45, nElo: 87.71 +/- 34.05
STC
Elo: 50.74 +/- 17.33, nElo: 101.17 +/- 34.05
LTC

[Sammy]

Hello everyone,

I'd like to announce the first public release of Hypernova, a UCI chess engine with an NNUE evaluation trained only on its own self-play games.

Download (Windows, source and network):
https://github.com/samarthkalani0-lab/h ... s/tag/v1.0

Executables

• Hypernova-1.0-avx2.exe - AVX2 CPUs (recommended)
• Hypernova-1.0-popcnt.exe - older 64-bit CPUs

The network is embedded in both executables.

UCI options: Hash (default 64 MB), Threads (default 1, Lazy SMP), EvalFile.
No pondering, no own book, no tablebase support.

Engine overview

• Magic bitboards
• PVS with aspiration windows, TT, null move, LMR (history-adjusted), LMP, futility and SEE pruning, singular extensions, killers, history and continuation history
• NNUE 768->256x2->1, trained with PyTorch on ~90 million positions from its own self-play (no other engine's data or evaluation)

Strength estimate: about 3140 on the CCRL 40/15 scale, from a 720-game gauntlet at 60+0.6 against 12 CCRL-rated engines (Loki 3.5.0 up to Igel 2.6.0; 95% interval 3107-3178). This is my own test, so the official CCRL result may differ.

Hypernova is open source under GPL-3.0. I'd be grateful if CCRL testers would consider including it, and any feedback or bug reports are welcome (GitHub issues or this thread).

Samarth Kalani