Coco Chess Engine

Precision Calculation.
Ruthless Search.

Coco is an autonomous tournament-grade C++20 chess engine pairing custom 768→512×2 NNUE neural evaluation with razor-sharp alpha-beta search and Lazy SMP concurrency.

Conceived, architected, and maintained by NotKaede-11. Free and open source under GPL-3.0.

Coco Emblem
Coco 1.5.1
Architecture Telemetry
STANDALONE
Evaluation
768→512×2 NNUE
Tournament Rating
CCRL 3027 / 2986
Concurrency
Lazy SMP Multi-Core
Move Generator
Magic Bitboards
CCRL 40/15 Rating
3027
Tested on official CCRL 40/15 list (+50/−50 over 98 tournament games; CCRL Blitz: 2986 Elo).
Throughput Speed
18M+
Nodes per second on AVX2 SIMD hardware with zero allocation inner loops.
Neural Architecture
512×2
Custom accumulator network embedded directly into every released binary.
Endgame Precision
6-Piece
Full Syzygy tablebase support with WDL and DTZ probe caching.
01 // CORE ARCHITECTURE

Engineered for depth & speed.

Every module in Coco is designed in modern C++20 with cache locality, zero dynamic allocation during search, and vector SIMD intrinsics.

ARCH // 01 EVALUATION

768→512×2 NNUE Network

A dual-perspective neural network architecture transforming board squares into dual 512-dimension accumulators with clipped ReLU activations. Incrementally updated on every move.

  • Accumulator16-bit Integer Quantized
  • SIMD AccelerationAVX2 & AVX-512 Intrinsics
  • DeliveryEmbedded in Executable
ARCH // 02 SEARCH DYNAMICS

Alpha-Beta & PVS Framework

Principal Variation Search with iterative deepening, adaptive aspiration windows, Late Move Reductions (LMR), Singular Extensions, and Reverse Futility Pruning.

  • SelectivityNMP, RFP, SEE Pruning
  • Move OrderingTT Hash, MVV-LVA, Killers
  • History TablesContinuation & Counter-Move
ARCH // 03 REPRESENTATION

Magic Bitboard Movegen

64-bit Bitboard representations for all piece topologies. Sliding attack masks for rooks and bishops are computed in O(1) time using precomputed Magic Hash tables.

  • Move GenerationPseudo-legal with Filter
  • Slider LookupsFancy Magic Bitboards
  • Zobrist Hashing64-bit Polyglot Compatible
ARCH // 04 CONCURRENCY & ENDGAME

Lazy SMP & Syzygy Tablebases

Multi-threaded scaling via lockless shared Transposition Tables across all worker threads. Seamless endgame resolution via Syzygy 3-4-5-6 piece WDL and DTZ probing.

  • MultithreadingLazy SMP (Shared TT)
  • TablebasesSyzygy WDL & DTZ
  • ProtocolUniversal Chess Interface (UCI)
02 // EMBEDDED NEURAL NETWORK

Zero external weights. NNUE baked inside.

Every Coco release binary includes the complete 768→512×2 neural network compiled directly into the executable. There are no separate .nnue weight files to download, locate, or configure.

PLUG & PLAY
Immediate Neural Evaluation

Load the Coco binary directly into Cutechess, BanksiaGUI, Nibbler, or Arena. The engine immediately boots into full NNUE mode with zero missing-file errors.

HCE SEPARATION
No Silent HCE Fallback

Unlike engines that silently degrade to Handcrafted Evaluation (HCE) when missing weight files, Coco guarantees 100% neural evaluation by default in every tournament game.

OPTIONAL OVERRIDES
Custom Network Testing

Engine researchers and neural network trainers can still benchmark custom experimental weights anytime via the standard UCI EvalFile option.

03 // INSTALLATION & RELEASES

Get Coco 1.5.1 for your platform.

Precompiled standalone binaries ready to be loaded into Cutechess, BanksiaGUI, Nibbler, Arena, or any standard UCI chess interface.

coco-chess-windows-x86-64-popcnt.exe
Windows 10/11 64-bit · Intel Core or AMD Ryzen with POPCNT / AVX2 support · Embedded NNUE
Embedded NNUE
AVX2 / POPCNT
Standalone Executable
Protocol UCI Standard
Evaluation 768→512×2 NNUE
Concurrency Lazy SMP Threads
Endgame Syzygy 6-Piece TB
UCI Quickstart Terminal
01 Launch Executable
$ ./coco-chess-windows-x86-64-popcnt.exe
02 Initialize & Configure UCI
> uci
> setoption name Hash value 256
> setoption name Threads value 4
> isready
03 Search Position
> position startpos moves e2e4 e7e5
> go movetime 1000
CREATOR

NotKaede-11

github.com/NotKaede-11

Coco Chess Engine was created and is independently developed by NotKaede-11 as an ambitious study in modern chess programming, low-latency SIMD evaluation, and autonomous neural training pipelines.

Named after Coco from Witch Hat Atelier—built with that same spirit of quiet discovery and disciplined craft. Coco Chess Engine and CuckooChess are different projects. The engine is developed completely in the open, validated through continuous CCRL tournament testing, and published under the GNU General Public License v3.0.

Creator: NotKaede-11 C++20 Standard Open Source (GPL-3.0) CCRL 40/15: 3027 Elo CCRL Blitz: 2986 Elo