v0.1.3

Trained on 200K game( >= 2400 Elo) using the new architect. The resulting model is much more bigger and takes longer to inference. On the bright side, it plays quite strong and no longer makes obvious blunder.

v0.1.2

200k game

v0.1.1

Old 100K model with value head.

v0.1.0

New policy model trained with 100.000 player games from Lichess (2025 - February) database. It plays reasonable strong and really fast (as no searching required)

12M

Trained with 12M positions. Still makes blunder but there're some pretty good moves during a game.

Neural v0.0.1

This model was trained on 3 millions stockfish evaluated position released by Lichess. Nonetheless, its strength is relatively weak and often makes blunder.