NONOGRAMS BENCHMARK · LOCAL, NO THINKING

Qwen3.8 27B

Alibaba · open weights · released 2026-08-14 · route qwen3.8:27b · Model profile · Chess results · Sudoku results · Mines results · FreeCell results · Starjar results · Nations Royale results · Wildhaven results · Hustle Sim results · Elemental Armory results · Micro Mouse results · Ravelin results · Minecraft results

Results

Score
1795%: 0 to 38 · 6 of 36 base puzzles so far
Puzzles solved
1 of 6
Cells that match the solution
20.1%
Rows and columns whose clue it satisfies
17.2%
Wrong cells / left cells unknown / no grid
2 / 1 / 2
Near misses (three cells or fewer wrong)
0
Average stated confidence
38%
Average confidence when it had not solved it
26%
Sure but wrong (90% or more, unsolved)
0
Calibration (Brier, lower is better)
0.085
API-equivalent cost per solved puzzle
n/a
Tokens per puzzle (in / out)
1,453 / 6,505
Reasoning tokens per puzzle
n/a
Retries for an unreadable reply
3
API-equivalent cost per puzzle
n/a
Seconds per puzzle
775.4
Replies cut off at the length limit
0.0%
Aborted on our side (not counted)
0

By tier

TierNeedsSolvedOutput tokensSeconds
5 by 5Line logic, a round or two1 of 12,709362.7
10 by 10Line logic, about 3 rounds0 of 18,215494.2
15 by 15Line logic, about 5 rounds0 of 19,4471,695.3
20 by 20Line logic, about 8 rounds0 of 16,116470.5
25 by 25Line logic, about 11 rounds0 of 14,7371,206.5
BeyondLine logic stalls; needs contradiction or search0 of 17,805423.4

Cells by the depth of logic they need

Each cell is tagged with the round of line logic that settles it. Beyond line logic means only contradiction or search does.

DepthCellsRight
First pass40326.3%
Rounds 2-361320.7%
Rounds 4-733226.8%
Round 8 on1590.0%
Beyond line logic930.0%

Every puzzle