Arkon Engine 0.5 Max

Arkon Engine 0.5 Max is a chess engine by AlpinaAI with 20,202,322 trainable parameters. It fine-tunes Arkon 0.1 Max and combines the learned move policy and position value with a timed alpha-beta search.

Play in your browser

Download the complete play package. Extract it and run these commands in the extracted folder with Python 3.10 or newer:

python -m pip install -r requirements.txt
python play.py --hybrid-search --search-seconds 2 --port 8801 --open

You play White and Arkon plays Black. If the browser does not open, visit http://127.0.0.1:8801/. If that port is occupied, choose another with --port. Keep the terminal open during play.

The ZIP includes the weights, browser board, piece graphics, and Python source. The engine runs on CPU. Once the dependencies are installed, you can play offline without Stockfish or a GPU.

Play in a UCI chess GUI

Set the UCI executable to your Python interpreter and pass the full path to arkon.py followed by uci. From the extracted folder the command is python arkon.py uci. The UCI interface uses the hybrid search and a fixed two-second search budget.

Training and evaluation

The training corpus contained 198,529 positions, including 111,656 newly labeled with Stockfish 18 at a budget of 32,000 search nodes per position. The split used 178,156 for training and 20,373 for validation. The published FP32 checkpoint is from epoch 4 of a nine-epoch run.

  • With the same short search, 0.5 Max won 9 of 16 games against 0.1 Max, drew 4, and lost 3.
  • In a separate match with 0.5 Max using hybrid search, it won 7 of 10 games against 0.1 Ultra, drew 2, and lost 1.
  • On 68 mate-in-two test positions, the 0.5 Max policy selected a correct first move 44 times. The 0.1 Max and Ultra policies scored 30 and 47.

See evaluation.json for game records and training-report.json for the training metrics. Get the standalone checkpoint from model.safetensors. The piece artwork license and attribution are in pieces.

Downloads last month

-

Downloads are not tracked for this model. How to track
Safetensors
Model size
20.2M params
Tensor type
F32
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for Convergent-ch/Arkon-Engine-0.5-Max

Finetuned
(1)
this model