Which Mac chip for local LLMs?

TL;DR

On Apple Silicon, two numbers decide everything: unified memory (GB) sets what fits, and memory bandwidth (GB/s) sets how fast it generates. Tiers rank roughly base < Pro < Max < Ultra, with two traps: the M3 Pro (150 GB/s) is slower on this axis than the older M2 Pro (200 GB/s), and there is no M4 Ultra — Apple's fastest, highest-capacity option is the M3 Ultra at 819 GB/s.

The two numbers

• Unified memory (GB) — shared between the CPU, GPU and the OS. This sets what fits: the model plus its KV cache must live here alongside everything else. Buy comfortably more than the model size and leave headroom (see the VRAM guide — on a Mac, unified memory is your VRAM).
• Memory bandwidth (GB/s) — sets how fast it decodes. Generating each token means streaming the model's weights through the compute units, so the width of the memory pipe is the ceiling on token speed (Bandwidth vs compute).

As a rough RAM-tier guide: an 8B model at 4-bit is happy in 16 GB; a 70B at 4-bit wants roughly 48–64 GB; large mixture-of-experts models need enough to hold all experts, not just the active ones (MoE vs dense).

Bandwidth by tier

The figures below are the catalog's own (see sources). Within a tier, more bandwidth means faster token generation for the same model.

Base chips are fine for small models. The Max tier is where Macs get genuinely quick for mid-size models. The Ultra tier is the top: the M3 Ultra reaches 819 GB/s and up to 512 GB of unified memory — the fastest and highest-capacity Apple option.

The Pro row hides the line-up's oddity: the M3 Pro is a regression. It dropped to 150 GB/s from the M2 Pro's 200 GB/s — Apple narrowed the memory bus that generation — before the M4 Pro recovered and passed both at 273 GB/s. If you are shopping used, an M3 Pro is slower for LLMs than an M2 Pro on this axis, despite being a generation newer.

The M4 Ultra trap

Apple's naming invites the assumption that every generation gets every tier. It does not.

Putting it together

For LLM speed the tiers rank roughly base < Pro < Max < Ultra, with the M3 Pro's 150 GB/s dip as the notable exception. In practice the buying decision runs memory-first: pick the unified-memory size that holds your target model with headroom, then take as much bandwidth as the budget allows within that. For per-chip fit and estimated speeds, see the Apple GPU pages.

Whichever chip you land on, Apple Silicon runs GGUF and MLX models well through Metal — see which runtime to use.

Sources & data

• Apple Silicon bandwidth figures (M3 100, M4 120, M2 Pro 200, M3 Pro 150, M4 Pro 273, M3 Max 400, M4 Max up to 546, M3 Ultra 819 GB/s) and the M3 Ultra's 512 GB memory ceiling: this catalog's GPU roster (DR-23-verified against manufacturer specs).
• RAM-tier fit guide (16 GB / 48–64 GB): rule of thumb — check exact fits in the calculator.