Eagle2.5-8B on DGX Spark — BF16, vLLM vision-language inference benchmark

Open this configuration in the VLM Inference Benchmark Explorer

Model	Eagle2.5-8B
Parameters	8.1B
Device	DGX Spark
Quantization	BF16
Image size	720p
Images per camera	1
Max Cameras (response ≤ 3 s)	8
Engine	vLLM

Cameras	Response (s)	TPS (tok/s)	Target
1	0.74	15.5	≤3 s
2	0.90	15.7	≤3 s
4	1.11	11.3	≤3 s
8	2.22	6.8	≤3 s
16	6.08	4.7	>3 s
32	13.4	3.8	>3 s
40	17.5	4.0	>3 s
44	18.9	3.8	>3 s
48	—	—	Failed

Images per request	Response (s)	TPS (tok/s)	Target
1	0.74	15.5	≤3 s
2	1.44	14.6	≤3 s
4	2.97	15.7	≤3 s
8	6.41	13.3	>3 s
16	14.8	13.1	>3 s
32	37.7	13.2	>3 s
64	109.2	9.0	>3 s
128	—	—	Failed

Eagle2.5-8B (8.1B parameters), served in BF16 format with vLLM on DGX Spark, starts answering a single camera that sends one 720p image after 0.74 s, then writes 15.5 tok/s.

With 64 images per camera at 720p, the answer starts after 109.19 s. With one 2K image, after 0.78 s.

At the default target of an answer starting within 3 s (720p, one image per camera), it keeps up with 8 cameras at once.