Qwen3-VL-4B-Instruct on DGX Spark — BF16, vLLM vision-language inference benchmark

Open this configuration in the VLM Inference Benchmark Explorer

Model	Qwen3-VL-4B-Instruct
Parameters	4.4B
Device	DGX Spark
Quantization	BF16
Image size	720p
Images per camera	1
Max Cameras (response ≤ 3 s)	32
Engine	vLLM

Cameras	Response (s)	TPS (tok/s)	Target
1	0.22	23.4	≤3 s
2	0.34	27.8	≤3 s
4	0.56	26.4	≤3 s
8	1.02	23.7	≤3 s
16	1.54	18.4	≤3 s
32	2.40	12.8	≤3 s
64	4.14	7.4	>3 s
128	8.42	3.8	>3 s
144	—	—	Failed

Images per request	Response (s)	TPS (tok/s)	Target
1	0.22	23.4	≤3 s
2	1.23	22.3	≤3 s
4	2.69	21.8	≤3 s
8	4.45	20.9	>3 s
16	9.25	19.3	>3 s
32	17.6	16.5	>3 s
64	25.6	12.7	>3 s
128	81.0	8.8	>3 s
256	279.6	5.5	>3 s
512	—	—	Failed

Qwen3-VL-4B-Instruct (4.4B parameters), served in BF16 format with vLLM on DGX Spark, starts answering a single camera that sends one 720p image after 0.22 s, then writes 23.4 tok/s.

With 256 images per camera at 720p, the answer starts after 279.62 s. With one 2K image, after 1.04 s.

At the default target of an answer starting within 3 s (720p, one image per camera), it keeps up with 32 cameras at once.