Qwen3-VL-30B-A3B-Instruct on Jetson AGX Thor — FP8, vLLM vision-language inference benchmark

Open this configuration in the VLM Inference Benchmark Explorer

Model	Qwen3-VL-30B-A3B-Instruct
Parameters	30B
Device	Jetson AGX Thor
Quantization	FP8
Image size	720p
Images per camera	1
Max Cameras (response ≤ 3 s)	8
Engine	vLLM

Cameras	Response (s)	TPS (tok/s)	Target
1	0.46	56.0	≤3 s
2	0.66	41.1	≤3 s
4	1.04	27.8	≤3 s
8	1.70	21.6	≤3 s
16	3.10	12.9	>3 s
32	3.94	9.0	>3 s
64	6.63	5.5	>3 s
96	9.81	3.8	>3 s
104	—	—	Failed

Images per request	Response (s)	TPS (tok/s)	Target
1	0.46	56.0	≤3 s
2	0.71	55.6	≤3 s
4	1.43	53.4	≤3 s
8	2.95	49.8	≤3 s
16	6.73	44.1	>3 s
32	17.3	35.2	>3 s
64	50.1	25.5	>3 s
128	160.3	16.5	>3 s
256	560.5	9.6	>3 s
512	—	—	Failed

Qwen3-VL-30B-A3B-Instruct (30B parameters), served in FP8 format with vLLM on Jetson AGX Thor, starts answering a single camera that sends one 720p image after 0.46 s, then writes 56 tok/s.

With 256 images per camera at 720p, the answer starts after 560.54 s. With one 2K image, after 2.07 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.