What your Mac can actually run

The question I hear most: will this run on my Mac? Good question — and the honest answer is a number, not a vibe.

Apple Silicon shares one pool of unified memory between CPU and GPU. A loaded model spends from that pool: its weights, plus the KV cache that grows with your context, plus working memory while it generates. Rule of thumb we use: keep total loaded models under about 75% of your RAM, and leave headroom for everything else.

That’s why Bedrova does a pre-load fit-check. Before a model loads, it adds up what you’ve already got resident and what this one will need, compares it to your free memory, and tells you plainly: it’ll fit, it’s tight, or it won’t — and if it won’t, it suggests a smaller quant or model that will. It loads, or it tells you why not. It doesn’t just crash.

And because we run several model types at once, we account for memory across all of them — your chat model, your GGUF text context, your image model, your search index worker, your perception model, your transcription model — as one budget.

Want specifics for your machine? The compatibility page lets you pick your RAM and see what fits.

— Rover