About

What this site is

Hands-on AI Lab publishes what happens when AI models are run on hardware that people actually own. Two machines do the work: a desktop with an RTX 4070 Ti and 12GB of VRAM, and an Apple M4 Pro laptop. Both are ordinary consumer hardware, and both hit limits that a rented A100 never does. Those limits are the subject.

The site covers two clusters of work. Local video generation came first, because ComfyUI and Wan2.2 on a 12GB card produce failures that are specific, reproducible, and badly documented elsewhere. Local language models came second. The two clusters share the same machines, so comparisons between a CUDA GPU and Apple Silicon hold the model, the quantization and the context length fixed.

What gets published

An article gets published when a measurement exists behind it. Every number on this site points to an experiment record with a fixed seed, the driver and runtime versions it ran against, and the date it was taken. Those records live in a public repository and feed the compatibility database directly. If a claim has no record and no source link behind it, the build fails before the page can go out.

Failed configurations are published alongside working ones. A list of settings that ran out of memory on a 12GB card is more useful to somebody stuck at 3am than another article explaining what VRAM is.

What does not get published

• Errors that could not be reproduced on hardware here.
• Benchmark numbers copied from someone else and presented as measurements.
• Articles generated and posted without a person reading the final text.
• Screenshots taken from other people's videos, blogs or broadcasts.

The rules behind those refusals are written down in the editorial policy and the AI policy, and the measurement procedure is in the methodology.

How the site is funded

Nothing on the site is sponsored, and no vendor has been given review or approval rights over anything published here. The plan is affiliate links on hardware recommendations and, later, display advertising. Neither is running yet. When affiliate links appear, they are marked at the point of use and explained on the affiliate disclosure page.

Corrections and contact

Measurements go out of date and mistakes get made. Both are handled in public through corrections. If a step in an article does not reproduce on your machine, that is worth reporting: write to contact@handsonailab.dev or open an issue on the repository. More ways to get in touch are on the contact page.