Triumviratus 8.1 20261028 - 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?

Triumviratus 8.1: pre-release (28 October 2026)

This is a pre-release for testing. bench 219346.

Strength

• Against 8.0 (official release): +14.5 ± 6.5 Elo over 2,728 games, SPRT [0, 5] passed. Both universal PGO
  builds, 8+0.08, 1 thread per engine, UHO openings.
• At 15+0.15 and 30+0.3: estimated at about +10 Elo, from shorter runs.
• The network alone: +10.3 ± 4.3 Elo over Consilium (the 8.0 network) with the same search, 30k nodes per move,
  8,332 games.

What is new

New network: Consilium Plebis

• Each side has its own expert, chosen by the non-pawn material that side is facing (8.0 chose one expert for both
  sides by the total piece count).
• Eight experts; the king-relative weights are int8, so the eight take the memory of four.
• A small pawn-king subnetwork, cached per thread, adds its view of the pawns and kings to every layer stack.
• Trained from 21 to 28 October 2026: pretraining of 496 epochs, finetune of 60 and a continuation of 3 down to a zero
  learning rate. The final network is the average of the last 3 epochs.
• Inference written for this network, faster than the first version with the same evaluations.

Search

• The search code was rewritten in the project's own form, block by block, each block first at the identical search
  tree.
• One table of piece values for the whole search (exchange evaluation, futility, capture order, threats), ready
  for a value per network expert; in this build the values are the same for every expert.
• Pruning margins per expert for the three experts with the most material (reverse futility, futility,
  reductions, null move), from an SPSA run.
• HistGrad: the quiet history learns more from a surprising best move and punishes little a move the ordering
  already rated low. +0.90 ± 3.41 over 11,166 games.
• No null move with a minor piece or less (NmpNpmFloor 3). +1.32 ± 2.29 over 6,844 endgame games.
• Every update of the move statistics and correction tables in one learning module (identical tree).

Fixes

• Zobrist keys from a 64-bit generator. The old keys had only 32 independent bits, which caused false hash hits:
  the 37.Re4 blunder of 8.0 in a CCRL game, reported by A Tin Man. +4.57 ± 3.56 over 9,660 games at 5+0.05.
• Race between threads in the transposition table, dead work in the evaluation, and a static exchange evaluation
  that ignored the gain of a promotion: found by Bastiii in a code review.
• go, d or eval before any position command no longer crashes: the engine starts from the initial position
  (issue #7).
• Tablebase wins print as cp 20000 minus the distance (GUIs show ±199.xx) instead of about ±94.
• Printed evaluation recalibrated for the new network: +1.00 again means a 50% chance of winning (fitted on
  183,630 positions of 8.1 in 3,588 games; with the 8.0 scale the 50% point was at +0.94). Display only: the search
  and the games do not change.

Speed

• Legal move generator ported from Zoomies and an AVX-VNNI variant, by A Tin Man (+6% to +9% on the Raptor Lake CPUs
  measured).

Thanks

• Maurizio Platino (MortalSin), the project's main tester, who also shares the cost of training the networks.
  Maurizio has been a fundamental asset for this project: without Maurizio, the fast development that started with
  version 4 would never have been possible.
• A Tin Man and Bastiii, for the code contribution and the code review in this release.
• CCRL and the computer chess community.

Triumviratus 8.1 20261028 download  mirror