Arkon-Engine-0.5-Max
a compact chess engine for CPU
2 min readupdated
Built by the AlpinaAI team, this 20.2M-parameter chess engine is fine-tuned from Arkon 0.1 Max. It pairs neural position evaluation and move selection with classical time-managed alpha-beta search (hybrid search). The model calculates chess moves to play full games against human opponents or other engines. Its main upgrade over earlier releases in the series is a noticeable jump in playing strength after fine-tuning on positions labeled by Stockfish 18. Built for chess enthusiasts and developers alike, it runs completely offline on a standard CPU with no GPU or external engines required.
Key features
- Runs locally on CPU with no GPU or Stockfish installation needed.
- Combines a neural move policy with time-managed alpha-beta search.
- Ships as a ready-to-run archive complete with weights, piece graphics, and a browser-based web board.
- Supports the standard UCI protocol for connecting to third-party chess GUIs.
What you can do
- Launch the local web UI and play against the engine's Black pieces right in your browser.
- Connect the engine via UCI to popular chess GUIs for game analysis or engine matches.
- Use the weight checkpoint (model.safetensors) in your own neural chess algorithm experiments.
Example
python -m pip install -r requirements.txt
python play.py --hybrid-search --search-seconds 2 --port 8801 --open
Read more
DetailsTraining: Trained on a corpus of 198,529 positions (including 111,656 new ones labeled by Stockfish 18 at a depth of 32,000 nodes); released as an FP32 checkpoint from epoch 4.
Matches: 9 wins, 4 draws, and 3 losses in 16 games against 0.1 Max; 7 wins, 2 draws, and 1 loss in 10 games against 0.1 Ultra.
Tactics: 44 correct first moves out of 68 in mate-in-two tests (up from 30 on 0.1 Max).
Requirements: Python 3.10 or newer, runs on CPU.










Comments