Pick a model size. See which Macs run it — and how fast.
Every Apple Silicon Mac below is checked against the model's real memory footprint, with a tokens-per-second estimate from its memory bandwidth and the cheapest memory configuration that fits.
Your model
How the fit — and the speed — are estimated
Speed. Generating each token means reading every active weight out of memory once, so on Apple Silicon decode speed is limited by memory bandwidth, not compute: tokens/sec ≈ bandwidth ÷ bytes read per token. Nobody hits the theoretical ceiling, so we discount it by what llama.cpp actually achieves — about 65% on base and Pro chips, 57% on Max chips, and 45% on Ultras, whose two stitched-together dies cost some efficiency. MLX is often 10–20% faster than that.
Mixture-of-experts. A mixture-of-experts model like Gemma 4 26B-A4B has to fit all 26B parameters in memory, but only reads the ~4B active ones per token — so it's sized like a 26B model and runs closer to a 4B one. Expert routing costs some efficiency, so we discount MoE speed a further 30%.
Context. The KV cache grows with every token of context, adding to memory and to what has to be read per token. The first speed number is for a short conversation; the second is with the context window full, which is the slowest it gets.
Fit. The model's footprint is its quantized weights, plus the KV cache for your chosen context, plus ~1 GB of runtime overhead. It has to fit in what macOS lets the GPU use — about two-thirds of memory on Macs with 36 GB or less, three-quarters above that — not the full amount on the spec sheet.
Not covered: prompt processing (time to first token), which depends on GPU compute rather than bandwidth and is where newer chips with Neural Accelerators pull far ahead. Long pasted documents will feel much slower on older and smaller chips than the numbers here suggest.
Prices are US Apple Store prices with base storage, as of September 2026. Previous-generation Macs are listed without prices since they're only sold used or refurbished.