About ModelRigs
ModelRigs exists to make one practical decision easier: what hardware should you buy to run the models you care about locally?
I started ModelRigs when I decided I wanted to buy hardware for running my own AI models locally. I knew the outcome I wanted, but not what to buy. It was surprisingly difficult to answer basic questions with confidence: which models would fit on which machines, how fast they would run, whether more memory mattered more than a faster GPU, and where the useful limit of each system really was.
The information existed, but it was scattered across specification sheets, model repositories, forum posts, benchmark threads, and product pages. File size was often treated as memory requirement. Measurements and estimates were mixed together. Advice that was correct for one quantization or runtime was repeated as if it applied to everything.
ModelRigs grew out of the comparison I wanted for myself: one place where hardware, model size, quantization, memory fit, performance, and cost could be evaluated together.
The goal is not to produce a giant list of products or model files. It is to help you make a defensible choice. That means keeping a curated catalogue of hardware and canonical model releases, showing which combinations fit, and separating real benchmark observations from calculated estimates.
Every number should tell you what kind of number it is. Manufacturer specifications, measured benchmarks, aggregated results, estimates, and site assumptions are labelled separately. When the evidence is thin, ModelRigs should say so plainly.
You can explore the hardware catalogue, browse local models, or use the configurator to turn your own budget and workloads into a shortlist. The full calculation rules are documented on the methodology page.
I now have my own Mac Studio, and I will be using it to make ModelRigs better. It gives me a consistent machine for running models myself, checking setup instructions, comparing runtimes and quantizations, and publishing repeatable measurements instead of relying only on third-party reports.
One machine cannot represent every local AI setup, but it can make the site more grounded. The lessons from those runs will help improve the benchmark process, catch unrealistic estimates, and make the practical guidance clearer for anyone deciding what to run or buy.