Local AI verification record — 2026-10-27
Status: observed on one existing installation; not a universal benchmark.
Environment
- Ubuntu 26.04.1 LTS, kernel 7.0.0-38-generic, x86_64
- Intel Core i5-7300HQ, 4 cores
- 7.1 GiB RAM; 4.0 GiB swap
- NVIDIA GeForce GTX 1050 Mobile, 2,048 MiB VRAM, driver 580.178.04
- Ollama 0.35.1; qwen2.5:1.5b, ID 65ec06548149, 986 MB
Read-only inventory commands
ollama --version
ollama ls
ollama ps
lscpu
free -h
nvidia-smi --query-gpu=name,driver_version,memory.total,memory.used --format=csv,noheaderTest request
Prompt: Return exactly five short bullet points explaining why source verification matters.
Both requests used temperature 0, seed 42, num_predict 80, num_ctx 2048, stream false, and keep_alive 0. CPU used num_gpu 0; the GPU request used num_gpu 99.
Observed results
| Mode | Load ms | Prompt tokens/ms | Output tokens/ms | Rate | Stop |
|---|---|---|---|---|---|
| CPU | 14,394.8 | 41 / 1,262.5 | 59 / 10,491.1 | 5.62 tok/s | stop |
| GPU request | 8,246.1 | 41 / 147.0 | 80 / 2,225.0 | 35.95 tok/s | length |
The CPU output contained five points. The GPU output ended during point three at the configured cap. Output wording differed despite the fixed seed and temperature.
Limits
- One run per mode; no warm/cold series.
- Active layer placement and peak RAM/VRAM were not sampled.
- Installation, update, model pull, model removal, Windows, and macOS were NOT_TESTED.
- qwen3:4b was present but NOT_TESTED.
- No quality, purchasing, or universal compatibility conclusion follows from this record.
Original source
The formatted view above is generated safely from the bundled Markdown file. The original remains available for reuse.
- README.md MARKDOWN