Qwen3.5-0.8B on DGX Spark — BF16, vLLM vision-language inference benchmark

Open this configuration in the VLM Inference Benchmark Explorer

Model	Qwen3.5-0.8B
Parameters	0.9B
Device	DGX Spark
Quantization	BF16
Image size	720p
Images per camera	1
Max Cameras (response ≤ 3 s)	128
Engine	vLLM

Cameras	Response (s)	TPS (tok/s)	Target
1	0.08	115.3	≤3 s
2	0.10	123.8	≤3 s
4	0.16	113.7	≤3 s
8	0.29	98.3	≤3 s
16	0.50	75.8	≤3 s
32	0.84	51.2	≤3 s
64	1.17	28.4	≤3 s
128	2.09	15.1	≤3 s
256	4.89	8.1	>3 s
384	13.5	7.0	>3 s
416	15.8	6.9	>3 s
448	—	—	Failed

Images per request	Response (s)	TPS (tok/s)	Target
1	0.08	115.3	≤3 s
2	0.14	114.7	≤3 s
4	0.26	113.3	≤3 s
8	0.51	110.7	≤3 s
16	1.04	107.2	≤3 s
32	2.19	99.1	≤3 s
64	4.92	86.9	>3 s
128	12.1	69.5	>3 s
256	—	—	Failed

Qwen3.5-0.8B (0.9B parameters), served in BF16 format with vLLM on DGX Spark, starts answering a single camera that sends one 720p image after 0.08 s, then writes 115.3 tok/s.

With 128 images per camera at 720p, the answer starts after 12.12 s. With one 2K image, after 0.34 s.

At the default target of an answer starting within 3 s (720p, one image per camera), it keeps up with 128 cameras at once.