Real benchmarks for personal AI on consumer hardware.
▾ scroll ▾
a €7 / month CPU VPS
Single-board cloud servers like Contabo’s CPU-LLM tier. No GPU, but enough RAM to fit mid-size models in software. Always-on, cheap, and good for what it can actually do.
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
Same generation, slightly newer machine. The numbers tell you where the 4 GB tier stops and the 6 GB tier starts — and what an A3B mixture-of-experts looks like on a consumer GPU.
what we measure
Cold start. Steady state. Concurrent users. Tool calls. Memory. Watts. Heat.
Every run we believe in publishes its raw JSONL, its tee’d log, a human-readable summary, and a metadata file with the model checkpoint and the harness git SHA. Clone it, re-run it, tell us where we got it wrong.