Qwen3-VL-8B-Instruct on DGX Spark — BF16, vLLM vision-language inference benchmark

Open this configuration in the VLM Inference Benchmark Explorer

Model	Qwen3-VL-8B-Instruct
Parameters	8.8B
Device	DGX Spark
Quantization	BF16
Image size	720p
Images per camera	1
Max Cameras (response ≤ 3 s)	16
Engine	vLLM

Cameras	Response (s)	TPS (tok/s)	Target
1	0.36	14.4	≤3 s
2	0.57	14.9	≤3 s
4	1.03	14.4	≤3 s
8	1.66	13.2	≤3 s
16	2.60	10.9	≤3 s
32	3.94	7.8	>3 s
64	6.72	5.0	>3 s
80	8.03	4.1	>3 s
88	8.91	3.8	>3 s
96	—	—	Failed

Images per request	Response (s)	TPS (tok/s)	Target
1	0.36	14.4	≤3 s
2	0.71	14.3	≤3 s
4	1.35	14.1	≤3 s
8	2.73	13.6	≤3 s
16	5.84	12.9	>3 s
32	13.4	11.6	>3 s
64	33.3	9.4	>3 s
128	96.1	7.1	>3 s
256	311.6	4.8	>3 s
512	—	—	Failed

Qwen3-VL-8B-Instruct (8.8B parameters), served in BF16 format with vLLM on DGX Spark, starts answering a single camera that sends one 720p image after 0.36 s, then writes 14.4 tok/s.

With 256 images per camera at 720p, the answer starts after 311.58 s. With one 2K image, after 1.68 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.