SUDOKU BENCHMARK · LOCAL, NO THINKING

Gemma 3 4B

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

Results

Score
095%: 0 to 16 · 15 of 42 grids so far
Grids solved
0 of 15
Blank cells filled correctly
9.6%
Wrong cells / empty cells / changed a given / no grid
0 / 0 / 13 / 2
Broke the rules (complete grid repeating a digit)
0
Near misses (three cells or fewer wrong)
0
Average stated confidence
96%
Average confidence when it had not solved it
96%
Sure but wrong (90% or more, unsolved)
15
Calibration (Brier, lower is better)
0.913
API-equivalent cost per grid
n/a
API-equivalent cost per solved grid
n/a
Seconds per grid
20
Tokens per grid (in / out)
390 / 172
Reasoning tokens per grid
n/a
Retries for an unreadable reply
2
Aborted on our side (not counted)
0

By tier

TierNeedsSolvedOutput tokensSeconds
CozySingles only0 of 314820
GentleScanning0 of 225613
SteadyLocked candidates0 of 217615
TrickyHidden pairs and triples0 of 220430
DeviousWings and fish0 of 213913
ExpertDeep wings and fish0 of 215428
BeyondNo human technique finishes it0 of 213919

Cells by the technique they needed

Each blank cell is tagged with the simplest technique that fills it, given everything a solver could fill before it.

TechniqueCellsFilled correctly
Last free cell16916.0%
Scanning19212.0%
Locked candidates884.5%
Hidden pairs and triples5213.5%
Wings and fish636.3%
Deep wings and fish5413.0%
Search817.4%

Every answer