Qwen3.5-35B-A3B on DGX Spark — FP8, vLLM vision-language inference benchmark

Open this configuration in the VLM Inference Benchmark Explorer

Model	Qwen3.5-35B-A3B
Parameters	35B
Device	DGX Spark
Quantization	FP8
Image size	720p
Images per camera	1
Max Cameras (response ≤ 3 s)	16
Engine	vLLM

Cameras	Response (s)	TPS (tok/s)	Target
1	0.34	53.8	≤3 s
2	0.57	42.9	≤3 s
4	0.98	31.4	≤3 s
8	1.55	20.2	≤3 s
16	2.14	14.4	≤3 s
32	3.29	9.0	>3 s
64	5.51	5.4	>3 s
96	7.88	4.0	>3 s
112	9.11	3.5	>3 s
120	—	—	Failed

Images per request	Response (s)	TPS (tok/s)	Target
1	0.34	53.8	≤3 s
2	0.53	53.6	≤3 s
4	1.07	52.9	≤3 s
8	2.11	51.8	≤3 s
16	4.18	49.9	>3 s
32	8.80	46.5	>3 s
64	19.4	41.7	>3 s
128	46.5	34.5	>3 s
256	—	—	Failed

Qwen3.5-35B-A3B (35B parameters), served in FP8 format with vLLM on DGX Spark, starts answering a single camera that sends one 720p image after 0.34 s, then writes 53.8 tok/s.

With 128 images per camera at 720p, the answer starts after 46.52 s. With one 2K image, after 1.36 s.

At the default target of an answer starting within 3 s (720p, one image per camera), it keeps up with 16 cameras at once.