Qwen3-VL-4B-Instruct on Jetson AGX Thor — BF16, vLLM vision-language inference benchmark

Open this configuration in the VLM Inference Benchmark Explorer

Model	Qwen3-VL-4B-Instruct
Parameters	4.4B
Device	Jetson AGX Thor
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	0.33	28.2	≤3 s
2	0.46	27.4	≤3 s
4	0.65	26.2	≤3 s
8	0.95	23.7	≤3 s
16	1.58	19.1	≤3 s
32	2.23	13.6	≤3 s
64	3.61	8.4	>3 s
128	7.40	4.3	>3 s
160	9.72	3.4	>3 s
176	—	—	Failed

Images per request	Response (s)	TPS (tok/s)	Target
1	0.33	28.2	≤3 s
2	0.49	27.8	≤3 s
4	0.84	26.8	≤3 s
8	1.79	25.4	≤3 s
16	4.32	23.3	>3 s
32	11.3	19.4	>3 s
64	34.3	14.6	>3 s
128	113.5	9.9	>3 s
256	407.7	6.0	>3 s
512	—	—	Failed

Qwen3-VL-4B-Instruct (4.4B parameters), served in BF16 format with vLLM on Jetson AGX Thor, starts answering a single camera that sends one 720p image after 0.33 s, then writes 28.2 tok/s.

With 256 images per camera at 720p, the answer starts after 407.74 s. With one 2K image, after 1.26 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.