PC Magik Lab

Local models, real GPUs.

Watch Series View Repo

Methodology

Same Prompt

Every test run uses the exact same prompt to ensure consistency across comparisons.

Same Rig

We use the same hardware configuration for every test in a series to eliminate hardware variance.

One Variable

Only one variable (e.g., quantization, model size) is changed at a time to isolate impact.

Reloaded Model

The model is fully reloaded before every run to prevent VRAM fragmentation or cache interference.

Fully Public

Every prompt and result file is published for complete transparency.

The Rig

  • GPU NVIDIA RTX 3090 (24 GB)
  • Model Server LM Studio

About

PC Magik Lab explores the intersection of local hardware and large language models.