Triumviratus 4.2 - Own Trained Net

Triumviratus 4.2 — first own-lineage network

The first own-network build of Triumviratus. It evaluates with our own rubicon NNUE, trained from scratch on T80 data — no Stockfish network is shipped. The classical search is unchanged from 4.1; this release is about the network.

What's in this release

Two builds (pick one for your CPU). Each archive contains the engine + the network + a short README.

Archive	CPU requirement
Triumviratus_4.2_avx512.zip	AVX-512 + BMI2 — Intel Ice Lake+/Skylake-X, AMD Zen 4+
Triumviratus_4.2_avx2.zip	AVX2 + BMI2 — Intel Haswell 2013+, AMD Zen+

BMI2 (PEXT) is required. If the AVX-512 build won't start, use the AVX2 one.

Setup

1. Unzip.
2. Keep nn-rubicon-v1.nnue in the same folder as the executable (it's already in the archive).
3. Add the .exe to your UCI GUI (Cute Chess, Banksia, Arena, …).

Notes for testers

• 100% own network. nn-rubicon-v1.nnue is our own NNUE (own lineage), not a Stockfish net. Big-net-only by default (UseSmallNet off).
• Strength vs 4.1: the own network costs about −39 Elo vs the Stockfish-net 4.1 at 12+0.12 — the deliberate, accepted price of a fully own-lineage build. Search and speed are unchanged.
• Eval retuned for the new net: the NNUE post-processing constants were re-tuned (SPSA) for rubicon, since a different network wants a different eval-search coupling.
• Built with clang-cl + ThinLTO + profile-guided optimization.

License

GPLv3. The NNUE inference code (sfnnue/) is derived from Stockfish (GPLv3) and the whole project is distributed under the GPLv3; the network weights are our own. Full source in this repository.

Triumviratus 4.1

Triumviratus 4.1 — Search co-tune

+35.0 ±17.9 Elo over 4.0 at 20+0.2 (LOS ~100%, fastchess, UHO_2024 +080/+099), SPRT-verified.

What changed

• Large SPSA co-tune (~10k iterations, 18 parameters): pruning margins, history weights, LMR + SF-style depth-pruning block.
• No new features vs 4.0 — same search architecture, same NNUE (nn-b1a57edbea57.nnue). Just better-tuned constants.

Binaries

file	target
Triumviratus_4.1_avx512.exe	Intel Ice Lake+ / AMD Zen4+ (AVX-512 + VNNI + BMI2)
Triumviratus_4.1_avx2.exe	Intel Haswell+ / AMD Zen+ (AVX2 + BMI2)

Both clang-cl + ThinLTO + IR-PGO.

Setup

Place nn-b1a57edbea57.nnue and nn-baff1ede1f90.nnue next to the exe.

Triumviratus 4.0

Open-source (GPLv3) UCI chess engine: an original classical alpha-beta search with Stockfish's NNUE evaluation.

What's new in 4.0

Co-tuned search vector (+29.5 Elo): 55 search/ordering/time-management parameters re-tuned together via a large SPSA run, validated by SPRT (same net both sides, 20+0.08, LOS 99.7%).
New 24-byte transposition table caching a 16-bit static eval (skips NNUE forward passes on TT hits).
+27.6 Elo cumulative vs 3.7 (build-neutral SPRT) before the co-tune stacks on top.

Downloads — BMI2 required:
Triumviratus_4.0_avx2.exe — Intel Haswell 2013+ / AMD Zen+
Triumviratus_4.0_avx512.exe — Intel Ice Lake+ / AMD Zen4+
Setup: place the NNUE nets next to the executable — nn-b1a57edbea57.nnue (big, default) and nn-baff1ede1f90.nnue (small). Optional in-house net nn-rubicon-v1.nnue selectable via the EvalFile UCI option.

License: GPLv3. The NNUE evaluation and default network are derived from Stockfish (GPLv3); the search, move generator, and engine architecture are original. Full source in this repository.

Release 3.3: Syzygy tablebases, illegal-position crash fix, LMR revert

Triumviratus 3.3 Hybrid

Triumviratus 3.3 is a UCI chess engine (~3450+ Elo, CCRL scale) combining alpha-beta search, NNUE evaluation, an experimental policy network, and ABDADA parallel SMP. This release adds Syzygy endgame tablebase support and fixes a crash on illegal positions.

Highlights

• Syzygy tablebases — perfect endgame play: WDL probing in search + DTZ probing at the root (converts wins / holds draws). Tables up to 5 men, via Fathom.
• Crash fix — eliminated an access violation (0xC0000005) on illegal / king-capture positions parsed from FEN.
• More robust search — history-based LMR reverted to the conservative ±1-ply clamp after match testing.

Full Changelog

• Add Syzygy tablebase probing (Fathom): in-search WDL + root DTZ, auto-loaded from a Syzygy/ folder next to the executable or via the SyzygyPath UCI option.
• Fix access violation on positions where the side-not-to-move is in check: guard td_negamax_abdada / td_quiescence against searching illegal king-capture positions.
• Revert combined butterfly+continuation-history LMR to butterfly-only ±1-ply clamp.
• Bump version to 3.3 (UCI id name + startup banner).
• Build output renamed to Triumviratus_3.3.exe.
• Updated README.md and FIXES_DOCUMENTATION.md.

Download & Setup

Place these files in the same directory as Triumviratus_3.3.exe:

• nn-b1a57edbea57.nnue (Big Net)
• nn-baff1ede1f90.nnue (Small Net)
• triumviratus_policy.bin (Policy Network weights)

Optional, for tablebase support: put your Syzygy .rtbw/.rtbz files in a folder and set the SyzygyPath UCI option to its absolute path (a Syzygy/ folder next to the exe is auto-detected). Setting SyzygyPath explicitly is recommended, since some GUIs launch the engine from a different working directory.

Recommended UCI options

• Hash: 1024 MB or more for long time controls.
• Threads: match your hardware.
• SyzygyPath: absolute path to your tablebase folder (empty = disabled).
• UsePolicy: false by default; true enables the experimental policy network.

Build

Visual Studio 2022 (toolset v143), Release | x64 — already configured for /O2, AVX2, intrinsics and LTCG:

& "C:\Program Files\Microsoft Visual Studio\2022\Community\MSBuild\Current\Bin\MSBuild.exe" `
  "Triumviratus_3.0.vcxproj" /t:Rebuild /p:Configuration=Release /p:Platform=x64

Requires an AVX2-capable CPU.

Known Issues

• A very rare (~1 in 400 games) illegal-move / disconnect anomaly has been observed in self-play and is under investigation; it affects both sides equally and does not bias match results. Tournament PGNs are being collected to reproduce it.

Triumviratus 2.0

Release of Triumviratus 2.0, a strong UCI Chess Engine.

Triumviratus 2.0 is a sophisticated UCI chess engine that employs advanced techniques, including bitboard representation and powerful algorithms such as alpha-beta pruning with Negamax, complemented by a quiescence search. The engine has evolved from its previous version, Triumviratus 1.0, which drew inspiration from "Vice" by Bluefever. Notably, version 1.0 lacked the bitboard representation that is now a key feature in the upgraded Triumviratus 2.0, enhancing its overall efficiency and capabilities in chess analysis and gameplay.