utMEASURED ON SCOUT’S TEST MACHINE
Measured speeds on a 12GB GPU.
Scout’s test machine downloads open models from the Ollama library and times them with the same settings. These numbers describe one machine and one setup, not every computer with the same graphics card.
- Graphics card
- NVIDIA GeForce RTX 3060 · 12 GB
- Processor
- Intel(R) Core(TM) i5-7600K CPU @ 3.80GHz
- System memory
- 30 GB usable
- Runtime
- Ollama 0.40.6 · NVIDIA driver 580.178.04
- Last updated
- Oct 20, 2026
2 measured models
Fastest generation first · not a quality ranking| Ollama model | Parameters | Weights | Writing (tokens/s) | Reading prompt (tokens/s) | In GPU memory | Tested |
|---|---|---|---|---|---|---|
| qwen3.5:0.8b ↗ Qwen3.5-0.8B profile → | 752.39M | Q8_0 · 1.2 GB | 132.2 | 8,581 | 100% | Oct 20, 2026 |
| gemma4:26b-a4b-it-qat ↗ gemma-4-26B-A4B-it-qat-q4_0 profile → | 25.2B | Q4_0 · 16.0 GB | 35.2 | 598 | 96% | Oct 20, 2026 |
How these speeds are measured
Each model is loaded once to warm up, then asked to write 512 tokens 3 times. Writing speed is the median of those runs. Reading speed comes from one prompt of about 1,500 tokens. All runs use temperature 0, a 4,096-token context, reasoning (“thinking”) turned off where the model supports it, and Ollama’s default settings, with nothing else using the graphics card.
“In GPU memory” is the share of the loaded model that fit in the graphics card’s memory. When it is below 100%, the rest runs from system memory and is much slower, so those speeds depend heavily on the processor and memory of this machine.
What these numbers don’t tell you
Speed is not quality: a fast model can give worse answers. Results change with the runtime version, driver, context length, longer conversations, and running several requests at once. Models that cannot generate text, and download sizes too large for this machine, are not tested. A model’s absence here is not a failed test.
Download and memory figures use decimal GB. See models that fit a 12GB budget · Read the memory guide