Qwen3.6-35B-A3B on DGX Spark — FP8, vLLM vision-language inference benchmark

Open this configuration in the VLM Inference Benchmark Explorer

Model	Qwen3.6-35B-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	0.40	51.0	≤3 s
2	0.69	40.4	≤3 s
4	1.15	29.1	≤3 s
8	1.83	19.7	≤3 s
16	2.63	12.7	≤3 s
32	3.89	7.9	>3 s
64	6.37	4.9	>3 s
96	8.59	3.6	>3 s
104	9.32	3.4	>3 s
112	—	—	Failed

Images per request	Response (s)	TPS (tok/s)	Target
1	0.40	51.0	≤3 s
2	0.62	50.8	≤3 s
4	1.23	50.4	≤3 s
8	2.46	49.6	≤3 s
16	4.85	48.2	>3 s
32	10.1	45.5	>3 s
64	22.1	40.7	>3 s
128	53.2	34.0	>3 s
256	139.9	25.6	>3 s
512	—	—	Failed

Qwen3.6-35B-A3B (35B parameters), served in FP8 format with vLLM on DGX Spark, starts answering a single camera that sends one 720p image after 0.4 s, then writes 51 tok/s.

With 256 images per camera at 720p, the answer starts after 139.86 s. With one 2K image, after 1.56 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.