Benchmark Zarum on your own hardware
Read three different timing measurements and compare your results with a Ryzen 7 7435HS and an RTX 4050 Laptop GPU.
A six-gigabyte laptop GPU in practice
The project ran a short Zarum 9B benchmark on a Ryzen 7 7435HS, 16 GB of RAM and an RTX 4050 Laptop GPU with 6 GB of VRAM. The figures describe that laptop and workload. CPU and CUDA used different output lengths.
| Mode | Measurement | Result |
|---|---|---|
| CUDA | Model load | 11.0 s |
| CUDA | First token after loading | 0.25 s |
| CUDA | Generation with 32 output tokens | 13.1 tokens/s |
| CPU, eight threads | Separate run with 24 output tokens | 6.8 tokens/s |
The first answer includes model loading and prompt processing. A following question in the same session can reuse a loaded model. Long chat histories and other applications using the GPU change the time you experience.
Measure your workstation
- Open Модели и настройки (model settings). Record the selected model, available system memory and VRAM, and the acceleration mode.
- Choose Измерить скорость на этом ПК (benchmark this PC). Record load time, time to first token and generation rate separately.
- Repeat with the same input and context. Then try a longer prompt representative of your work. Keep those timings alongside the short benchmark.
On a memory-constrained machine, the catalog includes Qwen3.5 2B and 4B. Zarum 9B and its vision projector together take 6.60 GB on disk. The running system also needs memory for context, buffers and Windows. Speech and image generation use separate engines and need their own measurements.