HARDWARE PICKER

What can I run?

Pick your hardware and we’ll sort every model with a memory estimate into what runs comfortably, what is tight, and what won’t fit.

Estimates come from observed file sizes and a fixed headroom rule, not benchmarks. How we estimate memory

ASSUMPTIONS

How these estimates work

Mac unified memory. We assume macOS lets the GPU use two-thirds of unified memory at 36 GB and below, and three-quarters above. The wired limit can be raised with sysctl iogpu.wired_limit_mb, which we do not assume.
GB vs GiB. The amounts you enter are marketing GB. Apple memory and GPU VRAM are binary, so we convert at 1 GiB = 1.073741824 GB — a “24 GB” Mac has about 25.8 decimal GB.
Headroom. A model needs its file size × 1.2 plus a 1.5 GB runtime allowance — the same formula the model pages use.
GPU offload. A GPU can spill into system RAM, so a model is tight when it fits with little VRAM headroom or when it fits VRAM plus half of system RAM (slower).
CPU only. CPU inference is much slower than a GPU, so we ask for more free RAM: a model runs well when its estimate is within 75% of system RAM, and is tight when the file plus the 1.5 GB runtime allowance fits within 90%.
Not listed. 31 models without a size estimate are not listed.

These are estimates from observed file sizes, not benchmarks. They don’t account for long contexts or the KV cache, and fitting does not establish runtime compatibility or speed.

What can I run?

Runs well

Tight

Won’t fit