AMD · Ryzen AI · 2024
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.
—
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 | ≤ 108 tok/s | |
| Qwen3.5 9B | Runs | ≤ 21 tok/s | |
| Gemma 4 12B | Runs | ≤ 17 tok/s | |
| gpt-oss-20b | Runs | ≤ 60 tok/s | |
| Qwen3.6 27B | Runs | ≤ 7 tok/s | |
| Qwen3.6 35B A3B | Runs | ≤ 65 tok/s | |
| Llama 3.3 70B | Runs | ≤ 2 tok/s | |
| gpt-oss-120b | Runs | ≤ 33 tok/s | |
| Qwen3.5 122B A10B | Runs, tight | ≤ 15 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.
LPDDR5X-7500 soldered: 120 GB/s, but the memory cannot be upgraded.
Two DDR5-5600 SO-DIMM slots (up to 96 GB): 89.6 GB/s. OCuLink port for an external graphics card.
A 5-bay NAS (diskless) with ECC DDR5 SO-DIMMs, 10 GbE + 5 GbE and an OCuLink port.
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 small always-on box or an all-in-one NAS for small models and light MoE models.
Yes if you want a small assistant box or a NAS that can also run models. For 70B-class models or faster answers, step up to Strix Halo.
llama.cpp with Vulkan (RADV) can use most of the system RAM: a 96 GB machine in the llama.cpp report loaded an 80B model at 8-bit (about 85 GB). We count 85% as usable. Memory speed depends on the box: soldered LPDDR5X-7500 gives 120 GB/s, DDR5-5600 SO-DIMMs 89.6 GB/s.
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.
| Model | Quant | Runtime | Generation | Prompt | Source |
|---|---|---|---|---|---|
| Qwen3-Coder-Next 80B | Q4_K_M | llama.cpp (Vulkan, RADV) | 15.3 tok/s | 57 tok/s | llama.cpp issue #19396 (ProgenyAlpha)Feb 10, 2026 |
| Qwen3-Next 80B | Q8_0 | llama.cpp (Vulkan, RADV) | 10.9 tok/s | — | llama.cpp issue #19396 (iz0eyj)Feb 6, 2026 |
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.
Two to three times the memory bandwidth, for a much higher price