Pelicans on a bicycle

Qwen3.8-Flash-Next on FNFL (main 7392490), split over two machines with two V100s each. Prompt: "Generate an SVG of a pelican riding a bicycle". Eight requests a model, temperature 1.0, normal thinking, no token cap. Every pelican finished on its own.

IQ3_XXSUD-Q4_K_XL
Pelicans drawn8 / 88 / 8
Requests at once48
Tokens generated, total213,098127,435
Tokens a pelican, median28,51615,300
Tokens a pelican, fewest – most8,332 – 47,66910,997 – 24,010
Decode speed a request, median133 tok/s81 tok/s
Time a pelican, median218 s190 s
Whole run447 s240 s
Throughput, all requests together477 tok/s530 tok/s

IQ3_XXS

8 / 8
pelicans drawn
213,098
tokens generated
28,516
median tokens a pelican
8,332 – 47,669
fewest – most tokens
133 tok/s
median speed a request (4 at once)
477 tok/s
throughput, all together
IQ3_XXS pelican 1
pelican_1
Tokens
13,342
Speed
132.3 tok/s
Time
101 s
Thinking
24,960 chars
Drawing
4,101 B SVG
IQ3_XXS pelican 2
pelican_2 · fewest tokens
Tokens
8,332
Speed
134.0 tok/s
Time
63 s
Thinking
14,637 chars
Drawing
3,275 B SVG
IQ3_XXS pelican 3
pelican_3
Tokens
40,040
Speed
128.4 tok/s
Time
312 s
Thinking
79,372 chars
Drawing
9,553 B SVG
IQ3_XXS pelican 4
pelican_4
Tokens
24,347
Speed
128.5 tok/s
Time
190 s
Thinking
42,027 chars
Drawing
9,526 B SVG
IQ3_XXS pelican 5
pelican_5
Tokens
32,685
Speed
127.7 tok/s
Time
256 s
Thinking
61,240 chars
Drawing
8,538 B SVG
IQ3_XXS pelican 6
pelican_6 · most tokens
Tokens
47,669
Speed
138.1 tok/s
Time
345 s
Thinking
95,890 chars
Drawing
10,063 B SVG
IQ3_XXS pelican 7
pelican_7
Tokens
33,377
Speed
135.8 tok/s
Time
246 s
Thinking
63,814 chars
Drawing
8,846 B SVG
IQ3_XXS pelican 8
pelican_8
Tokens
13,306
Speed
133.8 tok/s
Time
100 s
Thinking
28,923 chars
Drawing
3,245 B SVG

UD-Q4_K_XL

8 / 8
pelicans drawn
127,435
tokens generated
15,300
median tokens a pelican
10,997 – 24,010
fewest – most tokens
81 tok/s
median speed a request (8 at once)
530 tok/s
throughput, all together
UD-Q4_K_XL pelican 1
pelican_1 · fewest tokens
Tokens
10,997
Speed
76.5 tok/s
Time
144 s
Thinking
13,323 chars
Drawing
9,853 B SVG
UD-Q4_K_XL pelican 2
pelican_2
Tokens
13,693
Speed
79.3 tok/s
Time
174 s
Thinking
18,936 chars
Drawing
10,694 B SVG
UD-Q4_K_XL pelican 3
pelican_3
Tokens
11,435
Speed
76.5 tok/s
Time
150 s
Thinking
13,241 chars
Drawing
11,019 B SVG
UD-Q4_K_XL pelican 4
pelican_4
Tokens
16,037
Speed
81.8 tok/s
Time
196 s
Thinking
22,278 chars
Drawing
11,879 B SVG
UD-Q4_K_XL pelican 5
pelican_5
Tokens
14,562
Speed
79.8 tok/s
Time
183 s
Thinking
17,243 chars
Drawing
13,654 B SVG
UD-Q4_K_XL pelican 6
pelican_6
Tokens
16,137
Speed
82.6 tok/s
Time
196 s
Thinking
21,832 chars
Drawing
12,212 B SVG
UD-Q4_K_XL pelican 7
pelican_7 · most tokens
Tokens
24,010
Speed
100.3 tok/s
Time
240 s
Thinking
40,278 chars
Drawing
10,311 B SVG
UD-Q4_K_XL pelican 8
pelican_8
Tokens
20,564
Speed
94.1 tok/s
Time
219 s
Thinking
33,130 chars
Drawing
10,141 B SVG

Speed a request is the server's own decode rate. Throughput is all tokens over the whole run (first request's start to last request's end). The two runs had different numbers of requests at once, so their speeds are not like for like.