This test used a board image rendered from a known public FEN example, not a real user’s game photo. I tested it in TiinyOS with Qwen3.6-35B-A3B and reasoning on.
I wanted to know whether I could give TiinyOS a board image before asking about a chess position. I uploaded the PNG and asked for a square-by-square transcription and the piece-placement part of FEN, the code a chess program uses to record where pieces sit. I explicitly asked it to leave move history unknown.
The answer said there were 16 White pieces and 16 Black pieces. Then I checked the code. It marked the White pawn on e4 and the White knight on f3 as Black pieces. Seven more White pawns on the second rank were also written in lowercase, which FEN uses for Black. The piece counts and the code could not both be right.
Tiiny went further and filled in whose turn it was, castling rights and the move counters. Those fields depend on how the game reached this board; the image did not provide that record. It also skipped the square-by-square table I requested. Had I copied its full FEN into an analysis tool, I would have been asking about a different position. I did not run an engine in this test.
For another board photo, I would check the piece code against the image before asking for a move. This single test does not establish an accuracy rate.
No comments yet.