Qwen3.5-2B on DGX Spark — BF16, vLLM vision-language inference benchmark

Open this configuration in the VLM Inference Benchmark Explorer

Model	Qwen3.5-2B
Parameters	2.3B
Device	DGX Spark
Quantization	BF16
Image size	720p
Images per camera	1
Max Cameras (response ≤ 3 s)	64
Engine	vLLM

Cameras	Response (s)	TPS (tok/s)	Target
1	0.15	44.8	≤3 s
2	0.21	55.1	≤3 s
4	0.33	52.3	≤3 s
8	0.60	47.5	≤3 s
16	1.02	38.7	≤3 s
32	1.48	26.3	≤3 s
64	2.45	16.3	≤3 s
128	4.66	9.2	>3 s
192	6.97	6.5	>3 s
224	8.51	5.8	>3 s
240	8.91	5.4	>3 s
256	—	—	Failed

Images per request	Response (s)	TPS (tok/s)	Target
1	0.15	44.8	≤3 s
2	0.26	44.7	≤3 s
4	0.51	44.5	≤3 s
8	1.00	44.1	≤3 s
16	2.03	43.6	≤3 s
32	4.17	42.3	>3 s
64	8.93	39.8	>3 s
128	20.2	36.2	>3 s
256	—	—	Failed

Qwen3.5-2B (2.3B parameters), served in BF16 format with vLLM on DGX Spark, starts answering a single camera that sends one 720p image after 0.15 s, then writes 44.8 tok/s.

With 128 images per camera at 720p, the answer starts after 20.23 s. With one 2K image, after 0.69 s.

At the default target of an answer starting within 3 s (720p, one image per camera), it keeps up with 64 cameras at once.