With the CCRL Blitz and 40/15 tests of Triumviratus 8.0 still running, Maurizio Platino and I have trained the network that will power Triumviratus 8.1.
The idea behind 8.0, scaling the network's capacity with little loss of speed, is proving successful: in the current 40/15 test, Triumviratus 8.0 has played 332 games without a single loss (39 wins), with an estimated rating of 3649 (+-32).
The same idea guided the design of the 8.1 network, named Consilium Plebis. "Consilium" refers to a group of "experts", more precisely input buckets, each specialised in a different range of the opponent's non-pawn material and chosen separately for each side. "Plebis" refers to a small pawn-structure and king-position subnetwork, trained together with the rest and added to every layer stack. The details of the network are in the repositories: the architecture and its measurements in NETWORKS.md, and the network file, the training recipe and the charts of the run in Triumviratus-Networks.
Against the official 8.0, the pre-release scored +14.5 +- 6.5 Elo over 2,728 games at 8+0.08 (both universal builds, 1 thread), and about +10 Elo in shorter runs at 15+0.15 and 30+0.3.
I want to celebrate this pre-release by thanking Maurizio Platino (aka MortalSin) for an incredible contribution, both as the project's main tester and as a financial supporter, sharing with me the costs of training the networks on vast.ai.
The 8.1 pre-release is not meant for CCRL testing, as it is not the official and final release yet. It is meant for those who want to experiment with the engine under development and its new network.
I also want to take this moment to address some claims made by third parties about Triumviratus. Given its quick progress and its use of AI, it has been accused of being a Stockfish clone. I expected nothing different, as Stockfish has been the king of computer chess for many years. So I want to make clear why Triumviratus 8.1 is not a Stockfish clone:
- Network and inference: the network has its own architecture, and its inference has been rewritten from scratch.
- Search: the old search followed Stockfish 19 as its algorithmic base and depended heavily on its code; in 8.1 the search code has been rewritten in the project's own form.
- Tuning and constants: some search constants and values are still close to the ones Stockfish uses, but many of them will change with the next SPSA tuning runs.
- Novelties: many techniques were introduced by Triumviratus itself; they are listed on the GitHub page.
- Open source is meant to be a way to keep improving projects. With Stockfish being the state of the art for so many years, its ideas and implementations are bound to be reused and improved. No one can build a new top engine from nothing, skipping more than 15 years of research: we all stand on the shoulders of giants.
- Measured on the search code against Stockfish 19 (an analysis by a Claude Opus 5.5 agent): the lines identical or equal after renaming are 4.6%, between Stormphrax (2.8%) and PlentyChess (5.3%) measured the same way; the steps of the node search kept in Stockfish's order are 76%, within the 68-76% of Stormphrax, PlentyChess, Reckless and Viridithas. Many search constants still have Stockfish's values: the next SPSA tuning runs are meant to change that.
Once again, Maurizio has proved to be a fundamental asset for this project. Without Maurizio, the fast development that started with version 4 would never have been possible.
Thanks, Maurizio. Thanks to CCRL and to the computer chess community.
Francesco
Download the 8.1 pre-release
