/ 01
Search deeper
Iterative deepening, PVS-style search, aspiration windows, and quiescence search form the core. Transposition tables and history-based move ordering help focus the search.
OPEN SOURCE · C++20 · 4PC TEAMS
Four players.
One team instinct.
A different board demands a different engine. IronPhoenix brings team-aware search and native neural evaluation to four-player Teams chess.
Developed by Nick / KnightStackSys
THE FOUR-PLAYER FIELDTEAMS
14 × 14 cross board · 160 playable squares
4 playersTwo coordinated teams
C++20Native engine core
PhoenixNetPurpose-built NNUE
Open sourceBuilt to explore
01 / THE ENGINE
Search with the
whole team in mind.
IronPhoenix is built around the chess.com-style cross board. Its search and evaluation account for teammates, four-player rules, and terminal king captures.
/ 02
Explore alternatives
Principal variation reporting exposes the engine’s line. Use MultiPV to compare several candidate moves, with time or depth limits to control the search.
/ 03
Evaluate as a team
Handcrafted evaluation provides a deterministic fallback. PhoenixNet adds a native neural evaluator with separate own-king and partner-king feature streams.
02 / A DIFFERENT KIND OF CHESS
Your partner changes
the position.
Red and Yellow face Blue and Green. Players move in a fixed cycle, so the engine must read a position through the lens of both its own king and its partner’s.
IronPhoenix supports four-player castling, promotion, en passant, and FEN position setup. Positive evaluation scores favor the team whose player is currently moving.
Read the project overviewTHE TURN CYCLE
- RRedTeam Red + Yellow
- BBlueTeam Blue + Green
- YYellowTeam Red + Yellow
- GGreenTeam Blue + Green
Four turns. Two teams. Shared stakes.
03 / PHOENIXNET
Built for this board.
Trained for this game.
PhoenixNet is IronPhoenix’s native NNUE evaluator, designed specifically for four-player Teams chess.
The v1 architecture combines 65,280 sparse input features with 17 king buckets and a shared 128-wide feature transformer. Own-king and partner-king streams feed a 256 → 32 → 32 → 1 dense network.
Place ironphoenix.nnue beside the executable to load it automatically. Without a network file, the engine falls back to handcrafted evaluation.
FROM SELF-PLAY TO EVALUATION
01
Generate
Native self-play datasets with parallel workers and controlled MultiPV exploration.
02
Train
Train on the sparse features consumed by the engine. Append or resume dataset generation.
03
Export & search
Export the native NNUE format and load it alongside IronPhoenix.
256 → 32 → 32 → 1
04 / GET STARTED
Build. Search. Learn. Rise.
Bring a C++20 compiler, CMake 3.24+, and Threads support.
MSVC, Clang, and GCC are supported.
01 / Build from source
git clone https://github.com/KnightStackSys/IronPhoenix.git
cd IronPhoenix
cmake -S . -B build -DIRONPHOENIX_BUILD_DATASET_GENERATOR=ON
cmake --build build --config ReleaseThe Release configuration is used by multi-configuration generators such as Visual Studio. For a single-configuration generator, add -DCMAKE_BUILD_TYPE=Release to the configure command.
02 / Run a first search
uci
isready
setoption name Hash value 256
setoption name MultiPV value 4
position startpos
go depth 8Launch the built ironphoenix executable, then enter these UCI-style commands. Set MultiPV to 1 for the normal single-line search.