Caipora 3.8, an original UCI chess engine from Brazil, written from scratch in Rust with its own
NNUE. About 95 Elo stronger than 3.6 (the first public release) by our own measurements.
Strength
- Gauntlet at 8+0.08, 1 thread, 300 games per opponent, against Stash versions with known CCRL
Blitz ratings: ~3603 ± 25 CCRL Blitz estimated (Stash 33: 86.7%, Stash 37: 74.5%, all
600 games ended normally). This is our own estimate, not an official CCRL rating. - Lichess: caiporaBot, blitz 2893 on 2 October 2026.
Downloads
| File | For |
|---|---|
caipora-3.8-windows-x86-64-v3.exe |
Windows, CPUs with AVX2 (Intel Haswell 2013+, AMD Zen / Excavator+). Fastest. |
caipora-3.8-windows-x86-64.exe |
Windows, any x86-64 CPU (about 20% slower) |
caipora-3.8-linux-x86-64-v3 |
Linux, AVX2, static binary |
caipora-3.8-linux-x86-64 |
Linux, any x86-64 CPU, static binary |
The network is embedded; no extra files are needed. Bench: 5263766 for every build
(caipora bench). Checksums in SHA256SUMS.txt.
What is new since 3.6
Each change passed an SPRT on the clock (8+0.08, bounds [0, 5]) against 3.6, and each package was
confirmed against the previous version (+48.6 for 3.7 over 3.6, +9.5 for 3.8 over 3.7):
- NNUE accumulator cache per king bucket and fused updates: about 12% faster, same evaluation (+22).
- Quiet history bucketed by threats, continuation history four plies back, fail-low bonus (+20).
- Expected cut-node tracking and a richer LMR, also for captures and checks (+11).
- Capture history aware of threats on the destination square (+6).
- Non-pawn and continuation correction histories (+6).
- Time management also scaled by the best move's node fraction (+5).
Architecture unchanged otherwise: NNUE g6 (768x8 king buckets -> 1024)x2 -> 8 output buckets,
SCReLU, trained on Leela Chess Zero data (ODbL); PVS with singular extensions, Lazy SMP, pondering;
standard chess and Chess960 / DFRC.
UCI options
Hash (16 MB), Threads (1, up to 256), Move Overhead (10 ms), Ponder, UCI_Chess960,
EvalFile (<embedded>), Clear Hash.
AI usage
Caipora is written by Claude Code (Anthropic) under the direction, review and testing of its
author; every commit is co-authored by Claude. No code, constants or networks from other engines
are used. Full policy: AI_USAGE.md.
License: GPL-3.0-or-later.