Every test follows a strict, repeatable process so results are fair, transparent, and reproducible.
Every comparison uses identical prompts, ensuring a fair apples-to-apples evaluation across models.
All benchmarks run on a single, consistent hardware setup so the only thing that changes is the variable under test.
Only one parameter or model setting is altered per test, keeping results clean and interpretable.
The model is fully reloaded from disk before every single run to eliminate caching artifacts.
Every prompt and result file is published in the public repository for full transparency and reproducibility.
One machine. One GPU. No cloud. No shortcuts.
24 GB GDDR6X VRAM — the engine behind every benchmark in this series.
Single GPULocal model server used for all inference runs — consistent environment across every test.
Local ServerThis is a hands-on look at running AI locally — no marketing fluff, just real hardware and real tests.
Computer service and homelab operator from Poland. I run AI models on real hardware, publish every prompt and result, and share the findings through this YouTube series.
Watch episodes on YouTube and explore all prompts and results on GitHub.