The $899 Mac mini is not a local-AI machine. The $2,699 64GB panic buy is usually the wrong answer too.
This is a buying guide for the current Mac mini (M6 and M5 Pro) if you want to run open-weight models on your desk: which memory SKU to order, why Hugging Face file size is not RAM, and why people run out of unified memory three weeks later.
You will see:
- Why 16GB fails once macOS and the KV cache take their cut
- Why decode speed follows memory bandwidth (153 / 170 / 307 GB/s), not chip marketing
- Why Neural Accelerators change time-to-first-token (measured llama.cpp prefill 210 → 742 tok/s on a 9B Q4) without making the newer M6 faster at decode than the M5 Pro
- Dense 27B vs 35B MoE on the same bus (~14 vs ~70 tok/s, author-reported)
- The four-config ladder: skip 16GB, skip the awkward 24GB tier, buy 32GB M6 for most people, escalate to 64GB M5 Pro or Mac Studio only for 70B-class work
- When frontier cloud coding still wins, and why you should rent 27B and 70B on OpenRouter before you solder $1,400 of RAM
For ninety percent of local-AI buyers, the order is the M6 Mac mini with 32GB.
## Chapters
00:00 Two tabs, two answers
00:25 The $1,800 gap
00:51 A five-inch, four-watt machine
01:21 Unified memory vs PCIe
01:50 The prefill freeze
02:20 Neural Accelerators: 210 to 742
02:46 M6 vs M5 Pro
03:09 Decode is a memory pipe
03:47 Faster only if it fits
03:58 16GB files are not 16GB RAM
04:17 The KV cache bill
04:52 macOS takes a cut
05:21 Dense 27B taxes the bus
05:47 MoE: 70 vs 14 tok/s
06:30 Skip the $899 16GB trap
06:53 24GB is the awkward middle
07:15 Buy the 32GB M6
07:53 64GB, then Mac Studio
08:28 Cloud still wins frontier coding
08:56 Rent before you buy
09:11 Order 32GB
#macmini #localai #localllm #applesilicon #m6 #m5pro #llm #unifiedmemory #ollama #llamacpp #privateai #homelab #developer #artificialintelligence #mac #qwen
## User queries
mac mini
Best mac mini for local AI
Should I buy a Mac mini for local AI?
Is the $899 Mac mini enough for local LLMs?
How much RAM do I need on a Mac mini for local AI?
M6 vs M5 Pro Mac mini for running LLMs locally
Why does my local LLM run out of memory after a few weeks?
What is KV cache and why does it use so much RAM?
Can a 16GB Mac mini run 27B models?
Is 32GB or 64GB better for local AI on Mac mini?
Does the M5 Pro generate tokens faster than the M6?
Should I buy a Mac Studio instead of a Mac mini for local LLMs?
Mixture of Experts vs dense models on Apple Silicon
Is local AI better than Claude or GPT for coding?
How to test local models on OpenRouter before buying hardware
Mac mini memory bandwidth 153 vs 170 vs 307 GB/s
Why Hugging Face GGUF file size is not the RAM you need
## Sources
- Apple Newsroom Mac mini (2026-08) product photography and Neural Accelerator claims
- Apple Support HT203253 power table (M6 idle ~4 W) and Mac mini acoustics (~5 dBA idle)
- Hugging Face `unsloth/Qwen3.8-27B-GGUF`, file `Qwen3.8-27B-Q4_0.gguf` (~16.1 GB)
- Starry Hope, citing Alex Ziskind llama.cpp 9B Q4 prefill 210 vs 742 tok/s and separate 2.5s → 0.7s TTFT (author-reported; not Apple “up to 4×”)
- Loengnavy, Medium, M4 Pro same-hardware MoE vs dense decode (~78.8 vs ~14.7 tok/s; VO rounds to ~70 vs ~14) — author-reported
- AtomGradient, MLX inference bandwidth characterization on Apple Silicon
- OpenRouter live catch: `qwen/qwen3.8-27b` vs `meta-llama/llama-3.3-70b-instruct`, same `auth.ts` expiry-unit prompt
- NordicSilicon Mac mini LLM tables used only as secondary context; M6/M5 Pro tok/s rows are not treated as creator-measured
## Links