Same prompt, every run
The exact same input is sent each time — no variation in wording or length. Only one thing changes per test.
Local models, real GPUs.
// How we measure
No cherry-picked runs — just honest benchmarks under identical conditions, so every result is reproducible and trustworthy.
The exact same input is sent each time — no variation in wording or length. Only one thing changes per test.
All testing happens on a single machine — no cross-machine comparison. The only variable is what we're measuring.
The model is loaded fresh before every single test, eliminating cache or warm-up bias from the results.
Every prompt and every result file is published alongside the episode so anyone can verify or rerun the test.
// The rig
Everything runs on one machine — simple, repeatable, honest.
A single NVIDIA RTX 3090 with 24 GB of VRAM does all the heavy lifting. It's the only GPU on this rig, keeping every test consistent and easy to follow.
// About the series
YouTube Series — Running AI locally, on real hardware
PC Magik Lab is a YouTube series about running large language models locally on real, personal GPUs. No cloud APIs, no black boxes — just models executing on actual silicon in someone's room.
Each episode measures what it can measure: throughput, latency, quality at different quantizations and contexts — always with the same prompt on the same rig, with one variable changed per test. Every file is published so anyone can verify.
// Get involved
Every episode on YouTube, every prompt and result file in the repository.