Real benchmarks for personal AI on consumer hardware.
▾ scroll ▾
New · October 2026 · image, video and music generation
153 retained generated files, from a hundred-image prompt study to short music and imperfect videos. Prompts, timings, poor outputs and failed attempts are all included.
historical CPU-server tests
A CPU-only server can keep a small model available without a dedicated GPU. These are historical tests, with the hardware and settings recorded on each report. Check the test date before drawing conclusions about current hosting plans.
a 2019 gaming laptop · 4 GB GPU
An old HP gaming laptop with a four-gigabyte GTX 1050 — the smallest GPU we test that’s worth running language models on. Sub-billion-parameter models comfortably; small coding models with tuning; mixture-of-experts surprisingly far past the GPU’s VRAM ceiling.
a 2018 gaming laptop · 6 GB GPU
The Predator tests cover language models and media generation on a GTX 1060 with six gigabytes of VRAM. Configurations differ between campaigns; compare the model, quantization and memory offload before comparing speeds.
what we measure
Generation speed. Request time. Configuration. Evidence.
Read the recorded settings alongside each result. Download the retained data where available, check the sample size, and look for limitations before comparing two machines. Older archive pages have less complete evidence than the newer campaigns.
Apple Silicon · historical runs
The archive also includes M1 Mac tests using Ollama and MLX. Runtime, model size and available unified memory matter: open the individual report for the tested setup.
Browse the Mac records →