Tool
Can my Mac run it?
Unified memory decides what fits; memory bandwidth decides how fast it feels. Pick your machine and we will rank the models that actually run — no guesswork, no forum digging.
Recommended for your Mac
How this works
Models are matched to your machine by download size at 4-bit quantisation, with a headroom factor for macOS, the KV cache and your open apps. Speed is estimated from memory bandwidth — the number that actually predicts token throughput on Apple Silicon.
Every figure here is an estimate, not a benchmark run. The catalogue is reviewed as models are released; see the methodology.
Questions about the results
How is the tokens-per-second figure calculated?
It is an estimate: the chip's memory bandwidth (GB/s) divided by the model size on disk, multiplied by a fixed efficiency factor of 0.33. It assumes 4-bit weights, a short context and no other heavy apps. Treat it as a direction, not a benchmark — a real run with a long context will be slower.
Why does the tool use memory instead of GPU cores?
On Apple Silicon the GPU and CPU share one memory pool, so the hard limit is whether the weights and KV cache fit. Bandwidth then decides speed. GPU core count matters, but it is secondary to those two numbers for token generation.
What does "runs, tight" mean?
The model fits, but leaves little headroom once macOS, your open apps and a long context are counted. Expect swapping and a sharp slowdown if you push context length or keep a browser open. A smaller model usually feels better in that situation.
Does this work for image models like FLUX in ComfyUI?
Partially. Image models need noticeably more memory than text models of the same file size because of activations and the VAE. The tool applies a stricter budget for the image generation category, but ComfyUI itself still needs 16 GB or more to be comfortable with SDXL and 24–32 GB for FLUX-class models.