Qwen3.8-27B on DGX Spark — FP8, vLLM vision-language inference benchmark

Open this configuration in the VLM Inference Benchmark Explorer

Model	Qwen3.8-27B
Parameters	27B
Device	DGX Spark
Quantization	FP8
Image size	720p
Images per camera	1
Max Cameras (response ≤ 3 s)	2
Engine	vLLM

Cameras	Response (s)	TPS (tok/s)	Target
1	0.78	8.1	≤3 s
2	1.36	8.1	≤3 s
4	3.45	7.6	>3 s
8	5.26	6.6	>3 s
16	5.85	5.5	>3 s
32	7.88	4.0	>3 s
48	10.2	3.1	>3 s
52	—	—	Failed

Images per request	Response (s)	TPS (tok/s)	Target
1	0.78	8.1	≤3 s
2	1.12	8.1	≤3 s
4	2.25	8.1	≤3 s
8	4.56	8.0	>3 s
16	9.35	7.9	>3 s
32	19.8	7.6	>3 s
64	44.2	7.2	>3 s
128	107.3	6.4	>3 s
256	285.2	5.4	>3 s
512	—	—	Failed

Qwen3.8-27B (27B parameters), served in FP8 format with vLLM on DGX Spark, starts answering a single camera that sends one 720p image after 0.78 s, then writes 8.1 tok/s.

With 256 images per camera at 720p, the answer starts after 285.16 s. With one 2K image, after 4.2 s.

At the default target of an answer starting within 3 s (720p, one image per camera), it keeps up with 2 cameras at once.