Qwen3.5-27B on DGX Spark — FP8, vLLM vision-language inference benchmark

Open this configuration in the VLM Inference Benchmark Explorer

Model	Qwen3.5-27B
Parameters	27B
Device	DGX Spark
Quantization	FP8
Image size	720p
Images per camera	1
Max Cameras (response ≤ 3 s)	8
Engine	vLLM

Cameras	Response (s)	TPS (tok/s)	Target
1	0.54	8.1	≤3 s
2	0.96	8.1	≤3 s
4	1.75	7.7	≤3 s
8	2.65	7.0	≤3 s
16	4.28	5.8	>3 s
32	6.70	4.2	>3 s
48	9.10	3.3	>3 s
56	10.3	2.9	>3 s
60	—	—	Failed

Images per request	Response (s)	TPS (tok/s)	Target
1	0.54	8.1	≤3 s
2	1.04	8.0	≤3 s
4	2.03	8.0	≤3 s
8	4.12	8.0	>3 s
16	8.52	7.8	>3 s
32	18.0	7.6	>3 s
64	40.4	7.2	>3 s
128	97.8	6.5	>3 s
256	—	—	Failed

Qwen3.5-27B (27B parameters), served in FP8 format with vLLM on DGX Spark, starts answering a single camera that sends one 720p image after 0.54 s, then writes 8.1 tok/s.

With 128 images per camera at 720p, the answer starts after 97.84 s. With one 2K image, after 2.39 s.

At the default target of an answer starting within 3 s (720p, one image per camera), it keeps up with 8 cameras at once.