VLM Inference Benchmark Explorer
Qwen3-VL-30B-A3B-Instruct on DGX Spark — FP8, vLLM vision-language inference benchmark
Open this configuration in the VLM Inference Benchmark Explorer
| Model | Qwen3-VL-30B-A3B-Instruct |
|---|---|
| Parameters | 30B |
| Device | DGX Spark |
| Quantization | FP8 |
| Image size | 720p |
| Images per camera | 1 |
| Max Cameras (response ≤ 3 s) | 16 |
| Engine | vLLM |
| Cameras | Response (s) | TPS (tok/s) | Target |
|---|---|---|---|
| 1 | 0.36 | 50.4 | ≤3 s |
| 2 | 0.57 | 36.7 | ≤3 s |
| 4 | 0.95 | 24.6 | ≤3 s |
| 8 | 1.55 | 17.1 | ≤3 s |
| 16 | 2.29 | 11.9 | ≤3 s |
| 32 | 3.44 | 8.3 | >3 s |
| 64 | 5.63 | 5.1 | >3 s |
| 96 | 8.67 | 3.7 | >3 s |
| 104 | 9.23 | 3.4 | >3 s |
| 112 | — | — | Failed |
| Images per request | Response (s) | TPS (tok/s) | Target |
|---|---|---|---|
| 1 | 0.36 | 50.4 | ≤3 s |
| 2 | 0.58 | 49.9 | ≤3 s |
| 4 | 1.17 | 47.9 | ≤3 s |
| 8 | 2.43 | 45.0 | ≤3 s |
| 16 | 5.30 | 39.7 | >3 s |
| 32 | 12.9 | 32.4 | >3 s |
| 64 | 34.8 | 23.6 | >3 s |
| 128 | 110.1 | 15.1 | >3 s |
| 256 | 379.8 | 8.9 | >3 s |
| 512 | — | — | Failed |
Qwen3-VL-30B-A3B-Instruct (30B parameters), served in FP8 format with vLLM on DGX Spark, starts answering a single camera that sends one 720p image after 0.36 s, then writes 50.4 tok/s.
With 256 images per camera at 720p, the answer starts after 379.83 s. With one 2K image, after 1.5 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.
Qwen3-VL-30B-A3B-Instruct (30B parameters), served in FP8 format with vLLM on DGX Spark, starts answering a single camera that sends one 720p image after 0.36 s, then writes 50.4 tok/s.
With 256 images per camera at 720p, the answer starts after 379.83 s. With one 2K image, after 1.5 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.