New engine releases 2026 H2

[mysammoore]

SimpleChess 3.5
Estimated Elo: 3475+ CCRL 40/15
Author Sam Moore
https://github.com/LinkLikesLattes/Simp ... tag/v3.5.0

[Damir]

Where are the executables?

[mysammoore]

Where are the executables?

All executables are now pre-compiled in releases! https://github.com/LinkLikesLattes/Simp ... tag/v3.5.0

I did not realize people wanted/needed compiled binaries, I'm so used to handling my own compiles....

[Damir]

Thanks.

[Dariusz]

Hello,

I would like to introduce Chessenger, a UCI chess engine in C99 with a classical evaluation (no neural network). Version 1.6 is out, together with all earlier public versions:

https://github.com/chessengeriadoteu/chessenger

- Releases: v0.2 to v1.6 with binaries; v0.1 is Sedar, a program I wrote in 2002, source only.
- Binaries for macOS (Apple Silicon and Intel), Linux (x86-64 and ARM64) and Windows (x64 baseline, popcnt, avx2, bmi2, and ARM64).
- Single-threaded. UCI options: Hash (1 to 1024 MB), MultiPV, Clear Hash, Threads (1), Ponder. No own book.
- License: GPL-3. Code and values taken from other engines (Stash 37, OliThink 5.10.5, 4ku, PeSTO tables) are credited in the Thanks section of the README; the MIT licenses of OliThink and 4ku are in LICENSES/.

How it is made: the C code was written by AI coding assistants (mostly Anthropic's Claude) under my direction. I decide what to try and what stays in the engine; since August 2026 every change to search and evaluation is played out in engine games before it is kept. The README describes this in detail.

Strength: 1.6 measured 3146 [3131, 3162] on the scale of MCERL, the Mac Chess Engines Rating List I maintain (1500 games at 2 s + 0.2 s per move, Hash 128 MB, against ten engines, Ordo anchored to their MCERL ratings). MCERL ratings are not directly comparable with CCRL or CEGT.

Feedback and bug reports are welcome, on GitHub or here.

[NathanDrake]

Triumviratus 8.0: the first top engine with a mixture-of-experts network

https://github.com/Tors3/Triumviratus/releases/tag/v8.0

Testing Results (partial, matches still running; counted on complete opening pairs, 95% interval):

• vs Stockfish 19: -4.3 ± 6.3 (133+1, 240 games)
• vs Stockfish 18: +3.9 ± 7.5 (133+1, 180 games)
• vs Triumviratus 7.0: +129.9 ± 15.7 (20+0.2, 400 games)
• IQ4 suite (1 s, 1 thread, by Maurizio Platino): 155/183, Stockfish 19 142/183

Main Changes:

• Consilium, a mixture-of-experts network: four experts, one per phase of the game, only one active at a time
• PassedPawns, our own input block for passed pawns
• Restructured search with margins and evaluation corrections per network expert
• New time management: more time when the opponent surprises, less when behind on the clock
• Chess960 support and one executable for every x86-64 CPU

Thanks to Maurizio Platino, Mark Tang and Andres (EveAnn) for their testing (see Credits).

[looa]

Elephant Gambit 1.0.0 "Herd"
https://github.com/looak/elephant/releases/tag/1.0.0

The main addition is Lazy SMP. 4 threads against 0.15.2 on 1 thread scored +198 ±26 over 414 games (8+0.08s, 32MB hash, SPRT [0, 5]). On a single thread, it plays about the same as 0.15.2, because this release was mostly about correctness rather than strength.

What changed:
- New transposition table: 8 entries per 64-byte bucket, each a single atomic 64-bit word, so no torn entries across threads. Twice the entries per MB, and Hash now uses all the memory you give it (up to 32GB).
- Four hash bugs fixed: an en passant square hashed when no pawn could capture (missed threefolds), black-to-move FENs hashed without the side key, a non-pawn move onto the EP square flagged as en passant, and the fifty-move counter not resetting on pawn pushes.
- UCI fixes: ucinewgame clears the TT, go nodes/mate/searchmoves work properly, Threads 1-64.
- The engine now checks the incremental hash against a full recompute in testing, so this class of bug should get caught earlier.

Given the SMP gain, I'd be grateful if any testers would consider it for the 4-CPU list as well. As before, bug reports or anything odd in a GUI are very welcome.

Alex

[Chacal X]

Triumviratus 8.0: the first top engine with a mixture-of-experts network

https://github.com/Tors3/Triumviratus/releases/tag/v8.0

Testing Results (partial, matches still running; counted on complete opening pairs, 95% interval):

• vs Stockfish 19: -4.3 ± 6.3 (133+1, 240 games)
• vs Stockfish 18: +3.9 ± 7.5 (133+1, 180 games)
• vs Triumviratus 7.0: +129.9 ± 15.7 (20+0.2, 400 games)
• IQ4 suite (1 s, 1 thread, by Maurizio Platino): 155/183, Stockfish 19 142/183

Main Changes:

• Consilium, a mixture-of-experts network: four experts, one per phase of the game, only one active at a time
• PassedPawns, our own input block for passed pawns
• Restructured search with margins and evaluation corrections per network expert
• New time management: more time when the opponent surprises, less when behind on the clock
• Chess960 support and one executable for every x86-64 CPU

Thanks to Maurizio Platino, Mark Tang and Andres (EveAnn) for their testing (see Credits).

Misleading advertising; if it's a "Universal Compilation," it was supposed to work on SSE41 or AVX, but it doesn't!
Flavio/

[Modern Times]

Misleading advertising; if it's a "Universal Compilation," it was supposed to work on SSE41 or AVX, but it doesn't!
Flavio/

I think you misunderstood. AVX2 is a minumum requirement:

One executable for every x86-64 CPU with AVX2

Triumviratus_8.0_universal.exe the engine, for every x86-64 CPU with AVX2

[robertstepien]

We have to wait for the Maestro build.