Will it run? And how fast?

How the estimate works

Generating each token means reading the whole active model from memory once. So the ceiling is memory bandwidth ÷ active model size. Real machines hit a fraction of that ceiling; we calibrate that fraction against measured runs (tagged measured). Mixture-of-experts models only read their active experts per token, so they run far faster than their file size suggests, but still need the full file in memory. Macs can give the GPU roughly 75% of total RAM by default. Prices are approximate US list prices; check the current price before buying.

Add your machine's real numbers

Have Ollama installed? Run this and email the output to [email protected]. We add verified results to the table and tag them as measured.

ollama run qwen3:8b --verbose "Explain memory bandwidth in 200 words." 2>&1 | grep -E "eval rate|load duration"
sysctl -n machdep.cpu.brand_string 2>/dev/null || lscpu | grep "Model name"