Launching today
Mac LLM Calculator
Which Mac runs which LLM, and how fast
1 follower
Which Mac runs which LLM, and how fast
1 follower
Pick a Mac and a model, get two answers: does it fit in unified memory, and how many tokens per second to expect. Covers M1 to M6 plus the NVIDIA and AMD boxes people compare them against. The speed model and the measured data are open source.
Hi Product Hunt.
I rent Apple Silicon Macs for a living, which is exactly why this calculator exists in the shape it does. The question I get most is some version of "will a 70B run on my Mac", and the honest answer is two numbers that anyone can compute: whether the weights plus context fit in unified memory, and bandwidth divided by the size of the weights.
The speed part is where most calculators go wrong. Bandwidth alone predicts 525 tokens per second for a 3B model on an M5 Ultra, which is nonsense. Ours fits a second term for fixed per token overhead, and the two constants come from six models I measured on a base M4 through Ollama, median of three runs. The formula reproduces all six within 4%. Everything is in a repo under MIT, including the raw measurements under CC BY 4.0, so you can check the arithmetic rather than trust me: github.com/bagdaer1/mac-llm-calculator
And if the answer is that your own Mac is too small for the model you want, that is the gap my day job fills: macyou.co rents dedicated Mac minis and Mac Studios by the month, from $99, with Ollama and an OpenAI compatible API ready in about five minutes. The calculator gives the same numbers whether you buy, rent or already own the machine.
Two things I would love from this crowd: measured numbers on hardware I do not have, especially a 5090 or a DGX Spark, and a sanity check on the memory math for MoE models. Both go straight into the repo with attribution.