Per-chain output tokens, by model
summed over a chain's generations; from the inference setup eval
0
5k
10k
15k
20k
25k
30k
35k
40k
output tokens / chain
Qwen3-Coder-30B-A3B
Gemma4-26B-A4B
Qwen3.5-9B (thinking off)
<b>Qwen3-Coder-30B-A3B</b><br>n=22<br>median 20,197<br>IQR 12,052–26,248<br>range 2,207–38,768
<b>Gemma4-26B-A4B</b><br>n=15<br>median 12,328<br>IQR 5,712–18,566<br>range 1,461–40,799
<b>Qwen3.5-9B (thinking off)</b><br>n=29<br>median 11,045<br>IQR 5,900–17,524<br>range 1,065–27,818