Apple’s compact desktop, two live configurations on this site: the M6 base
model (16GB, 170 GB/s, 20W, $899, 2026) and the M5 Pro model (24GB, 307 GB/s,
30W, $1,699, 2026). Same 5-by-5-inch chassis Apple has used since 2010.
What it does well:
-
Entry price: $899 is the cheapest current-generation Apple route
into local inference, and it draws about as much power as a lightbulb.
-
M5 Pro headroom: 307 GB/s puts the Pro model in MacBook Pro 14 M5
Pro territory - 24GB handles 30B-class dense models at 4-bit or large
MoE models with tight quantization.
Where it falls short: 16GB on the base model limits dense models to
about 7-8B at 4-bit once KV cache is budgeted; memory is not upgradeable.
Run it locally: Ollama and LM Studio both run on the M6. The
modeldex results show what fits your exact
config and its estimated tok/s.