Gemma-4-31B-it on Jetson AGX Thor — NVFP4, vLLM vision-language inference benchmark

Open this configuration in the VLM Inference Benchmark Explorer

Model	Gemma-4-31B-it
Parameters	31B
Device	Jetson AGX Thor
Quantization	NVFP4
Image size	720p
Images per camera	1
Max Cameras (response ≤ 3 s)	8
Engine	vLLM

Cameras	Response (s)	TPS (tok/s)	Target
1	0.66	7.3	≤3 s
2	0.83	7.2	≤3 s
4	0.97	6.8	≤3 s
8	1.55	6.4	≤3 s
16	3.35	5.8	>3 s
32	4.92	4.7	>3 s
64	7.86	3.3	>3 s
80	9.69	2.9	>3 s
88	—	—	Failed

Images per request	Response (s)	TPS (tok/s)	Target
1	0.66	7.3	≤3 s
2	1.07	7.3	≤3 s
4	1.95	7.2	≤3 s
8	4.25	7.2	>3 s
16	10.0	7.2	>3 s
32	29.1	7.0	>3 s
64	118.2	7.0	>3 s
128	—	—	Failed

Gemma-4-31B-it (31B parameters), served in NVFP4 format with vLLM on Jetson AGX Thor, starts answering a single camera that sends one 720p image after 0.66 s, then writes 7.3 tok/s.

With 64 images per camera at 720p, the answer starts after 118.25 s. With one 2K image, after 0.72 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.