Mac mini 16GB to Studio 256GB: Which Mac Runs Your LLMs Best?

Key Principle: Capacity Sets the Boundary, Bandwidth Sets the Speed

The fundamental rule verified throughout this analysis:

Unified memory capacity (16/24/32/48/64/128/256 GB) decides whether a model fits at all.

Memory bandwidth (153 / 307 / 614 GB/s / 1.2 TB/s) decides how fast tokens are generated.

Because each generated token requires reading the entire model weights from memory, the bottleneck is never compute but the width of the path from memory to compute units. This is why Macs with seemingly modest GPU cores can run models that NVIDIA cards cannot fit: they win on unified memory capacity but lose on absolute bandwidth at equal capacity .

1. Configuration Landscape: New Mac mini Only Has M6 and M5 Pro

Figure: Overview of current Mac mini / Mac Studio mainstream configurations. Prices vary by storage and region; refer to Apple's official site.

A critical correction: the new Mac mini (on sale 2026-09-22) comes in only two chips — M6 (base) and M5 Pro (pro) . There is no "regular M5". Apple skipped M5 on the mini, jumping from M4 to M6, while naming the pro tier M5 Pro. So any "Mac mini M5" search results actually refer to M5 Pro.

Sorted by memory tier, the lineup becomes clear:

Mac mini M6 (16/24/32 GB, 153–170 GB/s): First 2 nm process, 12-core CPU + 12-core GPU, each GPU core adds a neural network accelerator. Official AI peak performance up to 4× M4, but for bandwidth-bound tasks like llama.cpp, actual throughput still depends on that 153–170 GB/s.

Mac mini M5 Pro (24/48/64 GB, 307 GB/s): Up to 18-core CPU + 20-core GPU, Thunderbolt 5. The 64 GB top config is the largest model the mini can handle.

Mac Studio M5 Max (48/64/128 GB, 614 GB/s): Desktop-class bandwidth ceiling; 128 GB comfortably hosts 70B models.

Mac Studio M5 Ultra (96/192/256 GB, 1.2 TB/s): Current local LLM king; 256 GB can even tackle 235B-class MoE. 512 GB version arrives late October.

In short: mini handles "cheap and sufficient"; Studio handles "fits larger models" . Previous-gen M4 (120 GB/s) / M4 Pro (273 GB/s) are cheaper and still solid value picks, but the new lineup has fully transitioned to M6 / M5 Pro — speed tables below include both generations.

2. Can It Fit? Match Your Memory Tier

Figure: Each memory tier's ability to hold a single model at Q4 quantization.

Plain-language breakdown per tier (note M6 maxes at 32 GB, M5 Pro at 64 GB):

16 GB (M6 base): 8B runs smoothly; 14B barely fits (very short context); 32B+ won't load. Good for "toying with small models".

24 GB (M6 / M5 Pro entry): 8B and 14B stable; 32B won't fit.

32 GB (M6 max): 14B comfortable; 32B barely (low context); 70B no.

48 GB (M5 Pro / M5 Max entry): 32B stable; 70B only barely (requires lower bits, truncated context).

64 GB (M5 Pro max): 70B runs at Q3 (6–8 tok/s) — the largest model the mini can touch.

128 GB (M5 Max): 70B at Q4 comfortable; reaches threshold for 120B dense and 100B+ MoE — the starting point for a "serious local LLM machine".

192 / 256 GB (M5 Ultra): 235B-class MoE fits comfortably; the "buy ahead for the next few years" tier.

Counter-intuitive note: 96 GB and 256 GB M5 Ultra share identical bandwidth (1.2 TB/s) . The only difference is capacity. Therefore, "for the same model that fits in 96 GB, speed on 96 GB and 256 GB is nearly identical" — 96 GB isn't a slower tier, it just holds one tier less.

3. Real-World Speed: Bandwidth Rules

Figure: Community-measured tok/s ranges (Ollama / llama.cpp Metal / MLX), single-stream, Q4. Numbers vary with context and load; for reference only.

Key readings (vs. previous-gen M4 120 GB/s, M4 Pro 273 GB/s; new M6 / M5 Pro bandwidth up ~30% and ~12% respectively):

Small models (7B/8B) : M6 ~50–60 tok/s (vs. M4 40–45), M5 Pro 65–72, M5 Ultra hits 140–160 . Daily chat/coding plenty fast.

Medium models (27B/32B) : M6 13–16, M5 Pro 17–21, M5 Max 24–27, M5 Ultra 45–50 . This is the tangible feel of 170 → 307 → 614 → 1200 GB/s bandwidth steps.

70B : The watershed. 64 GB M5 Pro manages only 6–8 tok/s at Q3 (barely fits), while M5 Max / M5 Ultra sustain 20–25 tok/s — nearly 3× faster because the former is "cramped" while the latter "resides comfortably".

100B+ MoE (e.g., 235B) : 128 GB M5 Max struggles at 8–10 tok/s (Q2), but M5 Ultra from 96 GB delivers 13–16 tok/s . MoE decoding activates only a fraction of parameters, so "larger models can actually run faster than some smaller dense models" — a hidden ace of Apple's unified memory.

Rule reinforced: The more headroom a model has in memory, the more bandwidth advantage materializes; when crammed in (via lower quantization / cut context), speed gets dragged down by capacity pressure.

4. Direct Recommendations by User Profile

Skip the tables, pick your persona:

Tight budget, only need 8B/14B chat locally : Mac mini M6 24 GB (avoid 16 GB). Starts ~¥6,999, silent and power-efficient. For cheaper, check clearance previous-gen M4 24 GB.

Main local coding + 32B, want value : Mac mini M5 Pro 48 GB . Starts ~¥12,999; 307 GB/s pushes 32B to 17–21 tok/s — the sweet spot for most.

Want 70B resident, a proper local LLM workstation : Mac Studio M5 Max 128 GB (or last-gen M4 Max 128 GB). 614 GB/s, 70B stable at 20–25 tok/s.

Need 100B+ MoE / 235B, future-proofing : Mac Studio M5 Ultra 96/192/256 GB . 96 GB already runs 235B Q4 (~14 tok/s); 256 GB more headroom.

Explicitly not recommended : ① 16 GB as primary AI machine; ② 64 GB expecting smooth 70B Q4 (only Q3 barely); ③ Buying Apple to beat same-price CUDA cards on single-model throughput — that's not its strength.

5. Closing Summary

Capacity defines the boundary; bandwidth defines the speed. The essence of the Mac lineup is using "unified memory" to bring models that don't fit on NVIDIA cards onto your desk, then leveraging "low power, zero noise, always resident" to drive usage cost to near zero. It doesn't chase single-model benchmark crushing of CUDA; it lets you "fit it, use it anytime, no fan noise, no power bill". Work backward from the largest model you want resident to determine memory, then spend remaining budget on bandwidth — this selection method won't steer you wrong.

Note: New Mac mini / Mac Studio go on sale 2026-09-22; M6 memory bandwidth officially rated 153 GB/s (some channels list 170 GB/s), article uses range; speeds are bandwidth-model estimates + early community measurements, not per-unit benchmarks. Mac mini only comes in M6 / M5 Pro — no "regular M5". Verify latest chip and memory options on Apple's official site before purchase.