Apple’s fanless ultraportable, 2026 generation on the M5: 13-inch from
$1,199 and 15-inch from $1,399, 16GB or 24GB unified memory, 153 GB/s
bandwidth, 20W chip. The 24GB option is the one to pick for local models -
16GB leaves too little room above the OS and KV cache for anything beyond
7-8B dense at 4-bit.
What it does well:
-
Silence: no fan at all, so inference is inaudible.
-
Portability at 24GB: 30B-class MoE models (Qwen-class at 3-bit)
run on a 1.2kg laptop.
Where it falls short: no active cooling means sustained generation
throttles earlier than a Pro; 153 GB/s is the slowest Apple silicon tier
on this site, and tok/s on large models reflects that.
Run it locally: Ollama or LM Studio. The
modeldex results show what fits and its
estimated tok/s.