MedGemma-1.5-4B-it on Jetson AGX Thor — 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	Jetson AGX Thor
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	0.27	28.9	≤3 s
2	0.36	28.2	≤3 s
4	0.46	26.8	≤3 s
8	0.97	25.7	≤3 s
16	2.37	21.8	≤3 s
32	3.07	16.9	>3 s
64	4.61	11.6	>3 s
128	6.73	6.5	>3 s
192	8.88	4.9	>3 s
208	—	—	Failed

Images per request	Response (s)	TPS (tok/s)	Target
1	0.27	28.9	≤3 s
2	0.42	28.7	≤3 s
4	0.84	29.1	≤3 s
8	1.60	28.1	≤3 s
16	3.36	27.8	>3 s
32	7.57	26.6	>3 s
64	22.4	23.6	>3 s
128	93.0	22.3	>3 s
256	528.6	19.2	>3 s
512	—	—	Failed

MedGemma-1.5-4B-it (4.3B parameters), served in BF16 format with vLLM on Jetson AGX Thor, starts answering a single camera that sends one 720p image after 0.27 s, then writes 28.9 tok/s.

With 256 images per camera at 720p, the answer starts after 528.6 s. With one 2K image, after 0.34 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.