Triumviratus 8.0 - new version strong chess engine


Triumviratus is a strong, UCI-compliant chess engine written in C++. 
Author: Francesco (fraTorsello)   Rating Chess Engines Diary CEDR=3849

Triumviratus what's new?

The official release of 8.0 (October 2026). It replaces the pre-releases of 9 to 13 October. bench 137419.

What is new

  • Consilium, a mixture-of-experts network. The largest layer, the king-relative HalfKA inputs, is split into
    four experts, one per material band (chosen by the number of pieces, on top of the usual king buckets); only one is
    active at a time, so the layer has four times the weights at almost the cost of one. Piece-count buckets are
    standard after the accumulator (output buckets, layer stacks); to our knowledge this is the first engine that uses
    them on the input layer. The experts are chosen by a fixed rule, not by a learned gate.
  • PassedPawns, an input block of our own for passed pawns, alongside the threat, king-relative and pawn-pair
    inputs.
  • A restructured search, re-tuned by SPSA on Consilium, with search margins and evaluation corrections per
    network expert, move ordering with chess knowledge (knight outposts, the capture of a bishop of the pair) and the
    causal reduction of quiet moves.
  • Time management: more time when the opponent does not play the expected reply (the surprise rule), less time
    when behind on the clock, rules for defending a worse position.
  • Faster code at an identical search tree, and buffered UCI input and output with a high-resolution clock (a
    defect present since 3.0, worth most at short time controls).
  • Chess960 support.
  • One executable for every x86-64 CPU with AVX2, with the network embedded.

The techniques of the engine, compared with the published work of other engines:
NOVELTIES.md.

Strength

opponentgamesconditions+W =D −LElo (Triumviratus 8.0)
Triumviratus 7.040020+0.2, UHO 2024, 1 thread+177 =189 −34+129.9 ± 15.7
Triumviratus 7.04,000 (final)5+0.05, UHO 2024, 1 thread+1773 =1864 −363+128.0 ± 5.6
Stockfish 19240133+1 (CCRL Blitz), TopGM openings, 1 thread+5 =227 −8−4.3 ± 6.3
Stockfish 18600 (final)133+1 (CCRL Blitz), TopGM openings, 1 thread+20 =568 −12+4.6 ± 5.0

Stockfish 18 and 7.0 at 5+0.05 are final; the other matches continue after the release and the numbers will be updated in the README. All results are counted on complete opening pairs only. ± is the 95% interval on the opening pairs (pentanomial).

For the analysis of hard positions there is a separate build,
Triumviratus 8.0 Position Solver.

Files

File
Triumviratus_8.0_windows_universal.exeWindows, one file for every x86-64 CPU with SSE4.1; the best variant (up to AVX-512) is chosen at start-up
Triumviratus_8.0_androidAndroid, arm64 (every 64-bit phone); run it from a GUI such as DroidFish or from Termux
Triumviratus_8.0_android-dotprodAndroid, arm64 with the dot-product instructions (phones from about 2018 on): faster with the network
Triumviratus_8.0_macos_armmacOS on Apple Silicon (M1 and later): NEON with the dot-product instructions, PGO
Triumviratus_8.0_macos_x86macOS on Intel (AVX2: Haswell, 2013, and later), PGO
Triumviratus_8.0_linux_avx2Linux x86-64 with AVX2 and BMI2, statically linked (any distribution), PGO
Triumviratus_8.0_linux_avx2-nopextthe same without PEXT, for AMD Zen 1, Zen+ and Zen 2
Triumviratus_8.0_linux_sse41Linux x86-64 without AVX2 (SSE4.1 and POPCNT, CPUs before 2013), slower
nn-consilium.nnuethe network, already embedded in every engine file; needed only to build from source
  • At start-up the engine runs the best of six builds for the CPU (SSE4.1 for CPUs without AVX2, AVX2 without PEXT,
    AVX2, AVX-512, AVX-512 VNNI, AVX-512 VBMI2). The engine name is Triumviratus - 8.0 on every CPU. To force a build, set the environment
    variable TRIUMV_ISA to sse41, avx2-nopext, avx2, avx512, vnni512 or avx512icl.
  • Checked before publishing: bench 137419 on the builds the test machine can run (sse41, avx2-nopext, avx2, avx512), the
    UCI options, and that the executable imports only KERNEL32.dll.
  • The Android files are built with the NDK (clang 18, -O3, NEON), without PGO; their bench should read 137419.
  • The macOS files are built on GitHub's Mac machines (Apple clang, PGO with the same training as the Windows file)
    and their bench (137419) is checked on a Mac by the build itself (.github/workflows/macos.yml). They are not
    signed: after downloading, run chmod +x on the file and xattr -d com.apple.quarantine on it before adding it to
    a GUI. Reports from Mac users are welcome.
  • The Linux files are built on GitHub's Linux machines with clang and PGO and checked with the bench there
    (.github/workflows/linux.yml). Run chmod +x on the file after downloading.
  • Building from source: build/bootstrap_release_laptop.ps1, or source/universal/build_universal.ps1 -Official;
    Android with make android / make android-dotprod and the Android NDK.

Triumviratus is an exciting and fast-evolving chess engine that, as 1.10.2026, reached a CEDR rating of 3849. Silnik is currently competing in the 3rd League and holds the top spot after completing 30% of the season’s matches.



Comments

Popular posts from this blog

Reckless 0.9.0 - new version super strong chess engine

Reckless 0.10.0 - new version super strong chess engine

We already have the new Fritz 20 program - what does it look like and how does it work?

Torch-4 UCI v 1.01 - new version chess engine