Qwen3.6-35B-A3B on DGX Spark — NVFP4, vLLM vision-language inference benchmark

Open this configuration in the VLM Inference Benchmark Explorer

Model	Qwen3.6-35B-A3B
Parameters	35B
Device	DGX Spark
Quantization	NVFP4
Image size	720p
Images per camera	1
Max Cameras (response ≤ 3 s)	16
Engine	vLLM

Cameras	Response (s)	TPS (tok/s)	Target
1	0.30	78.2	≤3 s
2	0.51	60.5	≤3 s
4	0.90	47.3	≤3 s
8	1.52	33.3	≤3 s
16	2.45	20.5	≤3 s
32	3.53	12.2	>3 s
64	5.37	6.7	>3 s
96	7.67	4.7	>3 s
112	8.76	4.1	>3 s
120	9.41	3.9	>3 s
128	—	—	Failed

Images per request	Response (s)	TPS (tok/s)	Target
1	0.30	78.2	≤3 s
2	0.52	77.7	≤3 s
4	1.03	76.7	≤3 s
8	2.04	75.0	≤3 s
16	4.14	71.8	>3 s
32	8.70	66.0	>3 s
64	19.1	56.5	>3 s
128	47.4	44.4	>3 s
256	128.0	31.2	>3 s
512	—	—	Failed

Qwen3.6-35B-A3B (35B parameters), served in NVFP4 format with vLLM on DGX Spark, starts answering a single camera that sends one 720p image after 0.3 s, then writes 78.2 tok/s.

With 256 images per camera at 720p, the answer starts after 127.97 s. With one 2K image, after 1.34 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.