New engine releases 2026 H2

Discussion of anything and everything relating to chess playing software and machines.

Moderator: Ras

User avatar
Gabor Szots
Posts: 1630
Joined: Sat Jul 21, 2018 7:43 am
Location: Budapest, Hungary
Full name: Gabor Szots

Re: New engine releases 2026 H2

Post by Gabor Szots »

WinPooh wrote: ↑Fri Nov 06, 2026 11:14 pm GreKo 2026.11

https://greko.su/greko-2026.11.zip

Fixing bugs, some changes in search
Estimated strength +100...150 Elo over GreKo 2021.12 (2800...2900 CCRL)
Hi Vladimir,

It exits on my i7 12700K. Is it an AVX512 compile?
Gabor Szots
CCRL testing group
User avatar
WinPooh
Posts: 286
Joined: Fri Mar 17, 2006 8:01 am
Location: Russia
Full name: Vladimir Medvedev

Re: New engine releases 2026 H2

Post by WinPooh »

Gabor Szots wrote: ↑Sat Nov 07, 2026 6:32 pm Hi Vladimir,

It exits on my i7 12700K. Is it an AVX512 compile?
Hi,
no, it doesn't use AVX512 instruction set. Binary is built with g++.exe (Rev5, Built by MSYS2 project) 16.1.0, with default settings from Makefile. You can also use .sln file for MSVC build.
User avatar
WinPooh
Posts: 286
Joined: Fri Mar 17, 2006 8:01 am
Location: Russia
Full name: Vladimir Medvedev

Re: New engine releases 2026 H2

Post by WinPooh »

Gabor Szots wrote: ↑Sat Nov 07, 2026 6:32 pm It exits on my i7 12700K. Is it an AVX512 compile?
Need some additional info to clarify,

1. What is your platform, Windows or Linux?
2. Does it immediately exits if run it in console, or after some time?
User avatar
Gabor Szots
Posts: 1630
Joined: Sat Jul 21, 2018 7:43 am
Location: Budapest, Hungary
Full name: Gabor Szots

Re: New engine releases 2026 H2

Post by Gabor Szots »

WinPooh wrote: ↑Sat Nov 07, 2026 7:48 pm
Gabor Szots wrote: ↑Sat Nov 07, 2026 6:32 pm It exits on my i7 12700K. Is it an AVX512 compile?
Need some additional info to clarify,

1. What is your platform, Windows or Linux?
2. Does it immediately exits if run it in console, or after some time?
I open a separate thread for this.
Gabor Szots
CCRL testing group
triangleangle
Posts: 1
Joined: Mon Feb 20, 2023 12:21 am
Full name: Brighten Zhang

Re: New engine releases 2026 H2

Post by triangleangle »

Hey everyone, I'm excited to share my program Moraband, it's written in C++ and communicates via UCI. I estimate it's around 2200 in Blitz in Standard. Moraband also supports Chess960.

I've verified compilation for both Apple silicon and Linux. Current version is 1.3.7!

https://github.com/bistromath25/moraban ... tag/v1.3.7

Looking forward to seeing how Moraband performs out in the wild! Thanks!
genio
Posts: 3
Joined: Mon Oct 26, 2026 11:13 am
Full name: Genio Mo

Re: New engine releases 2026 H2

Post by genio »

Arabica 0.4.7

https://github.com/geniomdev/arabica-ch ... tag/v0.4.7

Some search improves.

Estimated strength ~+30

Multithreaded search:
1 thread vs 2 threads = +85 ± 23
1 thread vs 4 threads = +158 ±35
1 thread vs 8 threads = +214 ±35
Villi14
Posts: 1
Joined: Tue Jul 28, 2026 2:04 pm
Full name: Oleksandr Velykotskyi

Re: New engine releases 2026 H2

Post by Villi14 »

Maharajah 1.3.2

https://github.com/Villi14/MaharajahCXX ... tag/v1.3.2

New open-source UCI engine (MIT), author Villi .
Derived from BBC 1.2 by Maksim Korzh, with a self-trained NNUE (own self-play
data only, net embedded in the binary), Lazy SMP and a C++23 rewrite.
Full credits in the README.

Windows builds: windows-x64-avx2 (Haswell / Ryzen or newer, fastest) and
windows-x64 (any 64-bit CPU, use it if the AVX2 one exits). No AVX-512 is used.
Also Linux and macOS builds.

Estimated strength: about 2975 CCRL Blitz (1 thread), from 400-game matches
against Stash 25.0 and 27.0 at 60+0.6 (STRENGTH.md).
Options: Hash up to 1024 MB, Threads, Skill Level (default 10 = full strength).

I would be glad if the CCRL testers could include it (Blitz and 40/15,
4 CPU too if possible). Thanks!
Calogero Zarba
Posts: 1
Joined: Thu Nov 12, 2026 4:20 pm
Full name: Calogero Zarba

Fattorini 2026.11.0, a new UCI engine in Rust

Post by Calogero Zarba »

Hello everyone,

I'd like to announce Fattorini, a new open-source UCI engine written in Rust. The first release, 2026.11.0, is out.

It is named after Antonio Fattorini, an Italian immigrant to Yorkshire whose family's business made the first chess clock in 1883 (Fattorini & Sons).
  • Download: Fattorini 2026.11.0, one Windows x64 .exe with the network embedded and no other files. It needs AVX2 (x86-64-v3).
  • Source: github.com/cigizeta/fattorini, AGPLv3.
  • Search: alpha-beta with the usual pruning and reductions, Lazy SMP, Syzygy tablebases at the root and in search, pondering.
  • Evaluation: an NNUE with piece-square inputs and simplified threat features, SCReLU, trained on positions from Leela Chess Zero's test91 training data (ODbL), converted by Fattorini's own code. The project's aim is networks trained on a limited compute budget.
  • Estimated strength: about 3250 on CCRL Blitz, one CPU, from 400-game matches against Weiss 1.3, Halogen 10 and Velvet 3.2.0 at a scaled 2'+1". The method is in the README.
  • Author: Calogero Zarba, Italy. The code is AI assisted.
Options: Hash (16 MB by default), Threads (1), MultiPV, Ponder, SyzygyPath and Move Overhead. No own book.

Feedback and test results are very welcome.

I would be glad if the CCRL testers could include it. Thanks!

Calogero Zarba
grounzero
Posts: 2
Joined: Wed Oct 21, 2026 1:27 pm
Full name: Michael Grounds

Re: New engine releases 2026 H2

Post by grounzero »

Cadence 0.6.0

Cadence is an open-source UCI chess engine, for standard chess, Chess960 and double Chess960, with an alpha-beta search and a hand-written evaluation. There is no network file: each binary runs on its own.
github.com/grounzero/cadence
It also plays on Lichess as the bot Mithandros.

0.6.0 is the last release with only a hand-written evaluation. From the next release, Cadence adds an NNUE trained purely on its own self-play, with the hand-written evaluation kept as an option.

Strength
About 2796 +/- 9 (95%) on the CCRL Blitz scale. This is my own estimate, calibrated against CCRL ratings, and not a CCRL rating: a gauntlet of 0.6.0-rc.3 at CCRL Blitz conditions, anchored to seven engines on that list (0.6.0 differs from rc.3 only by a draw-scoring fix that passed a non-regression test). An eighth, Reckless 0.4.0, was played and left out under an anchor-spread limit set before the run; with all eight the figure is 2806.
Method, engines and every game: strength page
Every SPRT, passed or failed: test log

Download
Cadence 0.6.0 release
Take the AVX2 build, for x86-64 machines from about 2015 on, and the generic build if it won't start. There are builds for Windows, Linux (x86-64 and aarch64) and macOS (Apple silicon and Intel), with SHA256SUMS and build attestations on the release page.
macOS: the binaries are not notarized. After downloading, run chmod +x <file>, then xattr -d com.apple.quarantine <file>, or build from source.

Or use an install script, which checks the binary's checksum and bench, and its build attestation when gh is installed.
Linux, macOS, Git Bash or WSL:

Code: Select all

curl -fsSLO https://github.com/grounzero/cadence/releases/latest/download/install.sh
bash install.sh
Windows PowerShell:

Code: Select all

Invoke-WebRequest https://github.com/grounzero/cadence/releases/latest/download/install.ps1 -OutFile install.ps1
powershell -ExecutionPolicy Bypass -File install.ps1
Settings for testers
Threads 1 and Ponder off, both the defaults; Hash as your tests use (16 MB by default); no book needed; SyzygyPath optional, empty by default, which turns probing off; everything else at its default. The UCI id is "Cadence 0.6.0", and its bench reads 34,686,410 nodes.

I'd be grateful if Cadence could be considered for testing on the CCRL lists.

Bug reports are welcome on GitHub issues.