Download: github.com/ComputecOrg/caipora/releases/tag/v3.8 (Windows and Linux, AVX2 and generic x86-64 builds; the network is embedded, no extra file needed)
Source: github.com/ComputecOrg/caipora (GPL-3.0-or-later)
Strength: our own gauntlet at 8+0.08, 1 thread, 300 games each against Stash 33 and Stash 37 (86.7% and 74.5%) puts it at about 3600 +/- 25 on the CCRL Blitz scale. That is only an estimate, and I would be glad to see a real one. On Lichess it plays as caiporaBot (blitz around 2890).
Engine:
- Search: PVS, aspiration windows, singular extensions with multi-cut, null move, LMR with expected cut-node tracking, LMP, futility and SEE pruning, threat-aware quiet and capture histories, continuation and correction histories, Lazy SMP, pondering.
- Evaluation: NNUE (768x8 mirrored king buckets -> 1024)x2 -> 8 material output buckets, SCReLU, trained with bullet on Leela Chess Zero data (ODbL). The first networks were trained on Caipora's own self-play. No third-party network is used, not even as a starting point.
- Standard chess and Chess960/DFRC. UCI options: Hash, Threads, Move Overhead, Ponder, UCI_Chess960, EvalFile.
Credits: bullet by Jamie Whiting (training); Leela Chess Zero data (ODbL) via Linrock's binpacks; fastchess and Stash (testing).
Feedback and bug reports are very welcome.
Matheus de Carvalho Jesus
