The 2025-2026 MacBook Pro generation, three silicon tiers: M5 (14-inch,
153 GB/s, 20W, from $1,599), M5 Pro (14-inch, 307 GB/s, 30W, from $1,999),
and M5 Max (16-inch, 614 GB/s, 60W, from $2,499 with 128GB at $4,499).
Battery life is unchanged by inference workloads being bursty: sustained
LLM decode pulls full power, so plan for the charger on long runs.
What it does well:
-
M5 Max 614 GB/s: a laptop that decodes large models at roughly half
a Mac Studio M5 Ultra’s rate while traveling with you.
-
M5 Pro sweet spot: 307 GB/s and up to 64GB covers 30-70B dense
models at 4-bit, and 8B models at high tok/s for coding assistants.
-
Battery-assisted portability: the only format on this site that
runs 100B-class MoE checkpoints away from a wall socket (M5 Max 128GB).
Where it falls short: thermals throttle sooner than desktops under
hour-long sustained generation; 20W on the base M5 is the entry tier.
Run it locally: Ollama, LM Studio, or mlx-lm. The
modeldex results rank every checkpoint by
fit and estimated tok/s on your exact memory config.