GaiaChess NNUE networks

Neural networks (NNUE) for the GaiaChess UCI chess engine.

Files are append-only: a network file, once published, is never modified β€” engine releases pin an exact file name.

Architectures

Two incompatible generations live here. Each network carries a 16-byte integrity footer (architecture hash + content hash) and the engine refuses, loudly, a network whose architecture is not the one it was built for β€” so a file from one generation cannot be fed to an engine built for the other.

GaiaNet T1 (gaianet-t1-*, gaianet-t2-*, gaianet-t3*) β€” 12 king buckets Γ— 768 PST inputs + filtered threat features (~41K inputs) β†’ 640 β†’ CReLU + pairwise β†’ 16 β†’ 32 β†’ 1, 8 output buckets.

GaiaNet T2b (gaianet-t2b-*) β€” the current generation. 12 king buckets Γ— 768 PST inputs (i16 rows) plus 64,368 auxiliary features in i8 rows: 4,560 pawn pairs and 59,808 filtered piece-interaction (threat) features, accumulated into a 1024-wide layer β†’ paired SCReLU β†’ 16 β†’ 32 β†’ 1, with 8 material-based output buckets and a PSQT head. 87 MB per file.

Trained with Bullet (T1) and the in-house trainer (T2b).

Networks

File Architecture Engine version Notes
gaianet-t1-1000.bin T1 never shipped archived for reproducibility: this network generated part of the training data used by its successors
gaianet-t2-1000.bin T1 4.2.0 – 4.2.5 previous default network
gaianet-t3b-1000.bin T1 unreleased +12.0 Β±9.5 Elo vs t2 (SPRT, 2588 games)
gaianet-t3e-1100.bin T1 unreleased t3b fine-tuned for 100 more superbatches at lr/10
gaianet-t3e-1100b.bin T1 unreleased last T1 default: same weights as t3e-1100 plus the 16-byte integrity footer
gaianet-t2b-e013.bin T2b unreleased previous default network. Fine-tune of the T2b line, 60 superbatches at lr 1e-4 cosine, WDL 0.4. +25.4 Β±9.7 Elo vs its parent (SPRT)
gaianet-t2b-e017.bin T2b unreleased (default in main) current default network. Same recipe and same parent checkpoint as e013, 60 superbatches at lr 1e-4 cosine, with WDL 0.3 instead of 0.4. +3.2 Β±2.6 Elo vs e013 (SPRT, 36 702 games)

Usage

Networks are embedded into the engine at build time:

curl -L -o nets/gaianet-t2b-e017.bin \
  "https://huggingface.co/jromanghf/gaiachess-networks/resolve/main/gaianet-t2b-e017.bin"
MODEL=nets/gaianet-t2b-e017.bin RUSTFLAGS="-C target-cpu=native" \
  cargo build --release --features nnue

Or loaded at runtime through UCI: setoption name EvalFile value <path>.

License

GPL-3.0-or-later, same as the engine.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support