YouTube Series — Benchmarks & Experiments

Local models, real GPUs.

A hands-on series testing and comparing AI language models on consumer hardware — one variable at a time, every result published.

Watch on YouTube Source & Results

Methodology

How we measure things

Same prompt, every time

The exact same input is fed to each model. No variation, no ambiguity — results are directly comparable.

Same rig, every run

All tests happen on a single machine. Hardware differences never confound the results.

One variable changed

Each test isolates a single factor — model, quantization, context window — so cause and effect are clear.

Model reloaded before every run

Cold starts, not cached inference. Every measurement reflects real-world conditions.

Fully reproducible

Every prompt and every result file is published open-source. Run the test yourself.

Setup

The hardware & software stack

RTX 3090 — 24 GB VRAM

NVIDIA GeForce RTX 3090

The GPU — 24 GB VRAM, GDDR6X

  • GPU NVIDIA RTX 3090 — 24 GB VRAM, GDDR6X memory bus
  • Model Server LM Studio — local inference engine for running models
  • Single rig All benchmarks run on one machine — consistent conditions every time

About

What is PC Magik Lab?

PC Magik Lab is a YouTube series dedicated to running AI language models locally on real consumer GPUs. No cloud APIs, no synthetic benchmarks — just hardware, open-source models, and honest measurements.

The goal is simple: figure out what runs well at home, share the results openly, and build a reference anyone can verify for themselves. Every prompt used in an episode and every result file is published to the public repository alongside each video.