MedGemma-1.5-4B-it on DGX Spark — BF16, vLLM vision-language inference benchmark

Open this configuration in the VLM Inference Benchmark Explorer

Model	MedGemma-1.5-4B-it
Parameters	4.3B
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	0.19	24.6	≤3 s
2	0.27	28.4	≤3 s
4	0.40	27.2	≤3 s
8	0.96	26.9	≤3 s
16	1.48	23.4	≤3 s
32	2.21	18.1	≤3 s
64	3.22	12.1	>3 s
128	5.70	7.6	>3 s
192	8.60	5.4	>3 s
224	13.7	5.4	>3 s
240	14.7	5.1	>3 s
256	—	—	Failed

Images per request	Response (s)	TPS (tok/s)	Target
1	0.19	24.6	≤3 s
2	0.33	24.5	≤3 s
4	0.62	24.5	≤3 s
8	1.24	24.2	≤3 s
16	2.52	24.4	≤3 s
32	5.47	23.7	>3 s
64	13.0	23.0	>3 s
128	43.5	21.5	>3 s
256	230.1	20.2	>3 s
512	—	—	Failed

MedGemma-1.5-4B-it (4.3B parameters), served in BF16 format with vLLM on DGX Spark, starts answering a single camera that sends one 720p image after 0.19 s, then writes 24.6 tok/s.

With 256 images per camera at 720p, the answer starts after 230.12 s. With one 2K image, after 0.23 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.