MiniCPM-V-4.6 on DGX Spark — BF16, vLLM vision-language inference benchmark

Open this configuration in the VLM Inference Benchmark Explorer

Model	MiniCPM-V-4.6
Parameters	1.3B
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	114.3	≤3 s
2	0.32	127.5	≤3 s
4	0.50	120.3	≤3 s
8	0.77	101.2	≤3 s
16	1.33	80.0	≤3 s
32	2.20	50.0	≤3 s
64	3.75	27.2	>3 s
128	5.51	12.9	>3 s
256	9.24	6.8	>3 s
384	21.5	5.5	>3 s
416	—	—	Failed

Images per request	Response (s)	TPS (tok/s)	Target
1	0.19	114.3	≤3 s
2	0.33	114.2	≤3 s
4	0.61	113.1	≤3 s
8	1.19	112.1	≤3 s
16	2.34	109.2	≤3 s
32	4.80	104.4	>3 s
64	9.67	96.1	>3 s
128	19.9	83.2	>3 s
256	—	—	Failed

MiniCPM-V-4.6 (1.3B 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 114.3 tok/s.

With 128 images per camera at 720p, the answer starts after 19.86 s. With one 2K image, after 0.29 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.