Qwen3.6-27B on DGX Spark — FP8, vLLM vision-language inference benchmark

Open this configuration in the VLM Inference Benchmark Explorer

Model	Qwen3.6-27B
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	0.55	8.1	≤3 s
2	0.96	8.1	≤3 s
4	1.75	7.8	≤3 s
8	2.65	7.1	≤3 s
16	4.27	5.8	>3 s
32	6.69	4.2	>3 s
48	9.09	3.3	>3 s
56	10.2	3.0	>3 s
60	—	—	Failed

Images per request	Response (s)	TPS (tok/s)	Target
1	0.55	8.1	≤3 s
2	1.04	8.1	≤3 s
4	2.03	8.1	≤3 s
8	4.13	8.0	>3 s
16	8.51	7.9	>3 s
32	18.0	7.6	>3 s
64	40.2	7.2	>3 s
128	97.4	6.5	>3 s
256	—	—	Failed

Qwen3.6-27B (27B parameters), served in FP8 format with vLLM on DGX Spark, starts answering a single camera that sends one 720p image after 0.55 s, then writes 8.1 tok/s.

With 128 images per camera at 720p, the answer starts after 97.36 s. With one 2K image, after 2.4 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.