Local models, real GPUs.
Every test follows the same rigorous protocol so results are comparable and reproducible.
Every model is evaluated against the exact same prompt inputs, ensuring a fair and consistent comparison across all tests.
All benchmarks run on a single, fixed hardware setup — one GPU, one system — so nothing changes except what we're testing.
Only one parameter or model changes per test. Everything else stays locked so the result reflects that single difference.
The model is fully reloaded from disk before every single run, eliminating any cached-state bias from previous tests.
Every prompt, configuration, and result file is published in the public repository so anyone can verify or reproduce the tests.
The entire methodology is open. No hidden parameters, no cherry-picked results — just honest benchmarks you can trust.
One rig. One GPU. No shortcuts.
24 GB GDDR6X — the workhorse powering every benchmark in the series.
Local model server used for inference — consistent, reliable, and open-source.
Everything runs on one machine. No cloud, no distributed inference — just one GPU doing the work.
NVIDIA GeForce — 24 GB VRAM
Who's behind PC Magik Lab and why.
Mateusz Piekut
Computer service operator and homelab enthusiast from Poland. This series is about running real AI models on real, local hardware — no cloud, no hype, just honest benchmarks you can reproduce yourself.
Watch the videos, dig into the data, or contribute.