Local models, real GPUs.
A YouTube series running AI inference on actual hardware — same prompt, same rig, one variable changed per test. Every result is reproducible and published.
Every test is designed so you can trust the numbers. Same prompt, same rig, one variable changed per test — and everything is published for inspection.
The exact same input is sent to every model. No paraphrasing, no tweaking — identical queries across all runs.
Every test runs on the same physical machine. No cloud instances, no different thermal headroom — one chassis, one set of conditions.
We change exactly one thing per test — model, quantization, temperature — so the signal is always clean.
The model is reloaded fresh before each test. No cached states, no lingering context — clean slate every time.
Every prompt and every result file is published in the repository — open for anyone to verify.
One machine. One GPU server.
A hands-on experiment in running large language models on consumer-grade hardware. No cloud APIs, no black boxes — just a GPU, some open-source models, and a lot of curiosity.
The goal is simple: figure out what's possible to run locally, how fast it actually goes, and publish every result so others can build on it.
Videos on YouTube, raw results and prompts in the open-source repo.