AMD · Ryzen AI · 2025
128 GB of unified memory in a mini PC for less than a Mac Studio: the homelab favourite for MoE models.
from $2,200 (typical 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 | ≤ 231 tok/s | |
| Qwen3.5 9B | Runs | ≤ 45 tok/s | |
| Gemma 4 12B | Runs | ≤ 36 tok/s | |
| gpt-oss-20b | Runs | ≤ 128 tok/s | |
| Qwen3.6 27B | Runs | ≤ 15 tok/s | |
| Qwen3.6 35B A3B | Runs | ≤ 139 tok/s | |
| Llama 3.3 70B | Runs, tight | ≤ 6 tok/s | |
| gpt-oss-120b | Runs | ≤ 93 tok/s | |
| Qwen3.5 122B A10B | Runs | ≤ 42 tok/s | |
| 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.
Sold by Framework (not on Amazon). Price after the 2026 memory-related increase.
A 5-bay NAS (diskless): five 3.5"/2.5" SATA bays plus five M.2 slots, 2×10 GbE, 250 W built-in power supply. Its LPDDR5X-8533 gives 273 GB/s, about 7% more than the 256 GB/s of the other Strix Halo machines.
Sold by Minisforum's own store (not on Amazon). Same NAS as the 64 GB model, LPDDR5X-8533 (273 GB/s).
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: gpt-oss-120b, Qwen3.6 35B and other MoE models in a small x86 box running Linux.
Yes with 128 GB, for MoE models. Prefer the Mac Studio M5 Max if you want more speed and no tinkering.
On Linux the GPU can use almost all of the 128 GB (GTT memory); on Windows the BIOS caps the dedicated share at 96 GB. Our figures assume about 90% is usable.
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.
Apple silicon
Largest popular model at 4-bit: Qwen3.5 122B A10B
128 GB at 614 GB/s: gpt-oss-120b and 122B MoE models at a good pace, almost silently.
Faster and silent, but pricier
DGX Spark / GB10
Largest popular model at 4-bit: Qwen3.5 122B A10B
128 GB with CUDA in a tiny box: the NVIDIA answer to Strix Halo and the Mac.
Same capacity with CUDA
Ryzen AI
Largest popular model at 4-bit: Qwen3.5 122B A10B
The mid-range Ryzen AI mini PCs and NAS: up to 96 GB of ordinary DDR5, fine for 8–30B models and small mixture-of-experts models, slow for big dense ones.
Much cheaper if you only run models up to about 14B