Apple · Apple silicon · 2026
Up to 64 GB at 307 GB/s in a quiet mini PC: 27–35B models at a comfortable pace.
from $1,699 (launch price)
Researched from manufacturer data and public sources; not tested by us.· Last reviewed Oct 11, 2026
Popular models at Q4_K_M with a 32K-token context. Memory needed from real model files; speed is the theoretical ceiling.
Model weightsContext cache (KV)RuntimeUsable limit (colour = verdict)
| Model | Verdict | Memory | Speed ceiling |
|---|---|---|---|
| Qwen3 1.7B | Runs | ≤ 277 tok/s | |
| Qwen3.5 9B | Runs | ≤ 54 tok/s | |
| Gemma 4 12B | Runs | ≤ 43 tok/s | |
| gpt-oss-20b | Runs | ≤ 153 tok/s | |
| Qwen3.6 27B | Runs | ≤ 18 tok/s | |
| Qwen3.6 35B A3B | Runs | ≤ 166 tok/s | |
| Llama 3.3 70B | Not enough memory | — | |
| gpt-oss-120b | Not enough memory | — | |
| Qwen3.5 122B A10B | Not enough memory | — | |
| Qwen3.5 397B A17B | Not enough memory | — |
Theoretical maximum for generating text: memory bandwidth divided by the bytes read per token. Real runs usually reach about half to three quarters of it.
Prices are editorial references (launch price and a typical street price on the date shown), not live prices. In 2026 the memory shortage has pushed many prices well above launch.
Prices are shown only if updated in the last 24 hours and may change; the retailer’s price at checkout applies. We may earn a commission from these links at no cost to you. Affiliate disclosure
Best for: A quiet home AI server for 27–35B models and MoE models up to about 45 GB.
Yes with 48–64 GB: the easiest quiet setup for mid-size models.
macOS lets the GPU use about two thirds of memory on machines with up to 36 GB, and about three quarters above that. The limit can be raised (sudo sysctl iogpu.wired_limit_mb=…), leaving less for macOS and other apps; our figures use the default.
Source: Manufacturer specifications
How well the popular local-AI tools run on it today (editorial assessment).
Measured results with their source. We never invent speeds.
No measured results yet. The table above shows the theoretical ceiling instead.
NVIDIA GPUs
Largest popular model at 4-bit: Qwen3.5 122B A10B
96 GB on one card at RTX 5090 speed: gpt-oss-120b and 122B MoE models entirely in GPU memory.
Ryzen AI
Largest popular model at 4-bit: Qwen3.5 122B A10B
128 GB of unified memory in a mini PC for less than a Mac Studio: the homelab favourite for MoE models.
AMD GPUs
Largest popular model at 4-bit: Qwen3.6 35B A3B
32 GB for about a third of an RTX 5090: the cheapest new 32 GB card.