# Ollama benchmark

- Server: `http://192.168.0.13:11434`
- Hardware mode: GPU offload (99 layers) (`num_gpu=99`)
- Think: auto (chat=off, curator=on)

| Model | Scenario | Runs | Prompt tok | TTFT p50 (ms) | Total p50 (ms) | tok/s p50 | Output tok (mean) |
|---|---|---|---|---|---|---|---|
| qwen3:8b | chat | 5 | 98 | 369 | 1301 | 23.8 | 25 |
| gemma2:9b | chat | 5 | 89 | 426 | 1681 | 22.1 | 24 |
| llama3.2:3b | chat | 5 | 102 | 400 | 1033 | 54.1 | 29 |
| mistral-nemo:12b | chat | 5 | 82 | 446 | 1451 | 19.7 | 18 |