Ryzen AI Max+ 395 (Strix Halo): 128GB of Unified Memory for Local LLMs
As of 2026-09-05. Benchmark numbers below are from the public strix-halo-llm-perf repo (llama.cpp, 128GB systems) — attributed, reproducible, and not our own runs.
TL;DR
- The Ryzen AI Max+ 395 is AMD's Strix Halo APU: 16 Zen 5 cores, a 40-CU Radeon 8060S iGPU, and 128GB of unified LPDDR5X — up to 96GB of it can be assigned to graphics.
- ~$2,500-class mini PCs give you the cheapest way to hold 96GB+ of model VRAM on one desk (a 397B-parameter repo exists on this hardware class).
- Real llama.cpp numbers: Qwen3 30B-A3B at 86.1 tok/s, GPT-OSS 120B at 53.4 tok/s, MiniMax M2.5 228.7B at 32.8 tok/s.
- Tradeoff: laptop-class memory bandwidth and a 40-CU iGPU — prompt processing is slow next to discrete cards; generation speed is the strong suit.
What Strix Halo actually is
Strix Halo is AMD's desktop-class APU platform: a 16-core Zen 5 CPU and a 40-CU Radeon 8060S iGPU sharing one pool of LPDDR5X. With 128GB of unified memory and AMD's variable graphics memory, up to 96GB can be handed to the GPU.
Mini-PC makers ship it in small desktops (GMKtec EVO-X2, Beelink GTR9 Pro) — a complete 128GB machine for roughly the price of a single mid-range workstation GPU. That is the pitch: VRAM capacity per dollar that discrete cards cannot touch.
Real benchmark numbers (llama.cpp, 128GB systems)
From the public strix-halo-llm-perf repo, best observed single-host results:
| Model | Params | Quant | Generation tok/s |
|---|---|---|---|
| Qwen3 30B-A3B MoE | 30.5B | Q4_K_M | 86.1 |
| Nemotron-3 Nano 30B-A3B | 30.5B | MXFP4 | 61.5 |
| GPT-OSS 120B | 120B | Q4_K_M | 53.4 |
| Qwen3-Coder-Next 80B-A3B | 80B | Q4_K_M | 42.7 |
| MiniMax M2.5 | 228.7B | Q3_K_M | 32.8 |
Read those rows again: a $2,500 mini PC generating from a 228.7B model at 32.8 tok/s, and a 120B at 53.4 tok/s. No discrete GPU can hold those models at this price at all.
Where it wins and where it loses
Wins:
- Biggest consumer VRAM pool per dollar — 96GB allocatable
- Runs 70B dense at high quants and 100B+ at Q3/Q4 that no single consumer card fits
- Whole-machine power ~85–120W profiles; silent desktop form factor
Loses:
- Memory bandwidth is laptop-class. LPDDR5X shared with the CPU means prompt processing (prefill) crawls next to a 3090's GDDR6X. Long contexts hurt.
- 40 CU of iGPU is no 5090: for small models a $400 used card out-generates it.
- Software stack: Vulkan or ROCm with llama.cpp; expect more setup fiddling than a CUDA card.
How to think about buying one
- You want one 70B–120B+ model resident, speed is secondary → Strix Halo is the cheapest path. Compare with a used A6000 48GB (~$2,200–2,800) which holds less but generates dense-70B faster — see our 48GB build guide.
- You want fast 7B–32B → a used RTX 3090 (~$1,300, 2026-09-28 snapshot) or two wins on speed; see the used GPU guide.
- You want 96GB+ on macOS instead → Mac Studio 192GB is the alternative at a much higher price; our Mac mini cluster guide covers the Apple route.
- Pair it with PAIR → a Strix Halo box and a MacBook are exactly the kind of mixed pool NVIDIA PAIR can schedule.
Check current tracked prices and $/GB on the value index, and model × hardware fits on the compatibility matrix.
Sources
- strix-halo-llm-perf — llama.cpp benchmarks on 128GB Ryzen AI Max+ 395 systems
- autoresearch — 397B-parameter runs on a $2,500 Strix Halo desktop
- AMD Strix Halo variable graphics memory coverage
- DiggTop data: value index · matrix · benchmarks
Third-party benchmark numbers are attributed to their repo and date; re-run on your own hardware before buying.