Mistral-Small-4-119B-2603 on DGX Spark — NVFP4, vLLM vision-language inference benchmark

Open this configuration in the VLM Inference Benchmark Explorer

Model	Mistral-Small-4-119B-2603
Parameters	119B
Device	DGX Spark
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	0.59	31.4	≤3 s
2	0.78	24.5	≤3 s
4	0.95	16.8	≤3 s
8	1.84	11.5	≤3 s
16	3.31	7.3	>3 s
32	5.64	4.7	>3 s
64	9.44	2.8	>3 s
80	10.0	2.4	>3 s
88	10.7	2.3	>3 s
96	—	—	Failed

Images per request	Response (s)	TPS (tok/s)	Target
1	0.59	31.4	≤3 s
2	0.93	31.5	≤3 s
4	1.60	31.1	≤3 s
8	3.10	30.7	>3 s
16	7.01	29.5	>3 s
32	16.9	28.2	>3 s
64	44.9	25.5	>3 s
128	—	—	Failed

Mistral-Small-4-119B-2603 (119B parameters), served in NVFP4 format with vLLM on DGX Spark, starts answering a single camera that sends one 720p image after 0.59 s, then writes 31.4 tok/s.

With 64 images per camera at 720p, the answer starts after 44.93 s. With one 2K image, after 0.83 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.