H2O-Lightning-4B on DGX Spark — BF16, vLLM vision-language inference benchmark

Open this configuration in the VLM Inference Benchmark Explorer

Model	H2O-Lightning-4B
Parameters	4.7B
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	0.21	21.2	≤3 s
2	0.28	25.3	≤3 s
4	0.49	23.8	≤3 s
8	0.89	22.1	≤3 s
16	1.43	17.9	≤3 s
32	2.24	12.4	≤3 s
64	3.86	7.8	>3 s
128	7.47	4.4	>3 s
144	8.47	4.0	>3 s
160	—	—	Failed

Images per request	Response (s)	TPS (tok/s)	Target
1	0.21	21.2	≤3 s
2	0.39	21.1	≤3 s
4	0.76	21.0	≤3 s
8	1.52	21.2	≤3 s
16	3.10	20.4	>3 s
32	6.56	19.6	>3 s
64	14.5	18.2	>3 s
128	34.8	16.0	>3 s
256	—	—	Failed

H2O-Lightning-4B (4.7B parameters), served in BF16 format with vLLM on DGX Spark, starts answering a single camera that sends one 720p image after 0.21 s, then writes 21.2 tok/s.

With 128 images per camera at 720p, the answer starts after 34.82 s. With one 2K image, after 0.95 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.