Arkon Engine 0.1 Max
Arkon Engine 0.1 Max is a neural chess engine by AlpinaAI. Its 20,202,322 parameters were trained from random initialization. The network predicts a move policy and a position value on a standard chessboard.
Play in your browser
Download the play package and extract it. Open a terminal in the extracted folder and run these commands with Python 3.10 or newer:
python -m pip install -r requirements.txt
python play.py --open
The board opens at http://127.0.0.1:8793/. You play White. Select a piece,
then select its destination. Keep the terminal open during play. Use
python play.py --port 8795 --open if the default port is occupied.
The ZIP includes the model weights, browser board and piece artwork. Local play runs on the CPU and works without Stockfish. Once Python dependencies are installed, you can play offline.
Use with a UCI chess GUI
Set the engine command to python arkon.py uci and run it from the extracted
folder. The UCI engine uses the same short search as the browser opponent.
To ask for a move from a position, run python arkon.py move "<FEN>".
Optional Stockfish duels
Install Stockfish and set the
ARKON_STOCKFISH environment variable to the full executable path before
starting the server. Open /?match=1 on the local board for an automatic
duel. With a CUDA-enabled PyTorch installation, /?match=max uses Arkon's
iterative deepening search and gives each engine five seconds per move by
default. Use --search-seconds to set another time limit.
Training and evaluation
The dataset contains 600,000 distinct positions, with 539,339 for training and 60,661 held out by game for validation. Training combined Stockfish 18 labels with generated mating positions. The model ran for 24 epochs. The released FP32 checkpoint comes from epoch 4, which had the lowest validation loss.
- Max won 22 games and drew 2 against Arkon 0.1 Pro across 12 paired openings with colors reversed and the same short search for both engines.
- The Max policy picked a winning first move in 97 of 113 mate-in-one positions and 21 of 37 mate-in-two positions. Pro scored 81 of 113 and 10 of 37, respectively. These are first-move policy scores.
These results are specific to this evaluation. See
evaluation.json for the games and
training-report.json for the per-epoch metrics.
The standalone weights are in
arkon-0.1-max.safetensors.