Gemma-4-26B-A4B-it on DGX Spark — BF16, vLLM vision-language inference benchmark

Open this configuration in the VLM Inference Benchmark Explorer

Model	Gemma-4-26B-A4B-it
Parameters	26B
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	0.41	24.0	≤3 s
2	0.50	21.1	≤3 s
4	0.63	15.8	≤3 s
8	1.27	12.2	≤3 s
16	2.41	9.2	≤3 s
32	3.95	6.8	>3 s
64	5.57	4.6	>3 s
96	8.27	3.5	>3 s
112	9.45	3.2	>3 s
120	9.42	3.1	>3 s
128	—	—	Failed

Images per request	Response (s)	TPS (tok/s)	Target
1	0.41	24.0	≤3 s
2	0.69	23.8	≤3 s
4	1.21	23.6	≤3 s
8	2.38	23.5	≤3 s
16	4.76	23.4	>3 s
32	9.53	23.0	>3 s
64	19.5	22.4	>3 s
128	40.0	21.5	>3 s
256	—	—	Failed

Gemma-4-26B-A4B-it (26B parameters), served in BF16 format with vLLM on DGX Spark, starts answering a single camera that sends one 720p image after 0.41 s, then writes 24 tok/s.

With 128 images per camera at 720p, the answer starts after 40.03 s. With one 2K image, after 0.44 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.