CHESS BENCHMARK · LOCAL, NO THINKING

Gemma 3 4B

Google · open weights · released 2025-03-12 · route gemma3:4b · Model profile

Opening gap: +44 (95%: -740 to +828). The interval includes zero, so we can't yet say whether it performs differently across the variants.

Standard

Rating
33295%: -243 to 907
Record
0 won, 0 drawn, 3 lost of 3
Accuracy (open positions)
82.1
Accuracy (all moves)
79.1
Average centipawn loss
87
Blunders / mistakes / inaccuracies per 100 moves
21.4 / 0.0 / 14.3
Matched Stockfish's first choice
21%
In Stockfish's top three
36%
Legal on the first try
64.7%
Illegal / invalid
15 of 29 answers
Forfeits
3
Adjudicated games
0
API-equivalent cost per game
n/a
API-equivalent cost per move
n/a
Seconds per move
11.9
Model minutes per game
1.1
Tokens per move (in / out)
381 / 18
Reasoning tokens per move
n/a
Aborted on our side (not counted)
0

By phase

PhaseMovesAccuracyCentipawn lossBlunders /100Engine match
Opening1482.18721.421%
Middlegame0n/an/an/an/a
Endgame0n/an/an/an/a

Chess960

Rating
28895%: -245 to 821
Record
0 won, 0 drawn, 3 lost of 3
Accuracy (open positions)
81.7
Accuracy (all moves)
81.7
Average centipawn loss
55
Blunders / mistakes / inaccuracies per 100 moves
6.7 / 6.7 / 33.3
Matched Stockfish's first choice
20%
In Stockfish's top three
33%
Legal on the first try
44.4%
Illegal / invalid
19 of 34 answers
Forfeits
3
Adjudicated games
0
API-equivalent cost per game
n/a
API-equivalent cost per move
n/a
Seconds per move
14.8
Model minutes per game
1.5
Tokens per move (in / out)
537 / 33
Reasoning tokens per move
n/a
Aborted on our side (not counted)
0

By phase

PhaseMovesAccuracyCentipawn lossBlunders /100Engine match
Opening1581.7556.720%
Middlegame0n/an/an/an/a
Endgame0n/an/an/an/a

Games