About Llama Builds

Running a language model at home comes down to a few hard numbers: how much memory a model needs, how much your GPU has, and how fast that memory is. Most buying advice still talks about gaming frame rates. We built Llama Builds to answer the questions that actually matter for local AI.

Why we started

Llama Builds began as Noah's notebook. He's a software engineer and founder who got hooked on local AI the first time he ran Stable Diffusion on his own RTX 3090. Running AI on your own machine, even when it isn't the state of the art, feels different.

Soon friends were asking him how to build their own machines. Some had a spare closet with a single outlet; others had 240-volt circuits and room for a rack. The information they needed was scattered across subreddits, forums, and X threads, and much of it was two years old, which is an eternity in local AI. So he started collecting builds he knew worked and organizing them by cost, memory, and the models they can run. That is still the core of the site.

What we publish

• Buying guides that pick specific GPUs, Macs, and mini PCs for each budget, and explain why.
• A hardware and model database covering 36 GPUs and systems and 28 open-weight models, with what runs on what, how much memory it needs, and roughly how fast it goes.
• Benchmarks from measured runs, every one linked to its source, alongside clearly labeled estimates.
• Real builds: 11 rigs from our own bench and the community, with full parts lists and prices.

How we stay independent

Llama Builds is reader-supported. When you buy something through our links we may earn a commission, at no cost to you. No company pays to be included in or to rank higher in our guides. Our picks come from the numbers: memory, bandwidth, price, and measured performance. Read our full affiliate disclosure.

How we estimate

We can't test every model on every GPU, so most speed figures on the site are estimates from a memory-bandwidth model that we calibrate against published measurements. Estimates are always marked with a tilde (~) or hatched bars, and measured numbers always cite their source. Our methodology page explains exactly how it works and where it can be wrong.

Get in touch

Built a rig you're proud of, spotted a mistake, or have a question we should answer? Message us on X (@AiFlux). We fix errors quickly and note significant corrections on the page.