64 GB: strong for 20B to 32B
A useful capacity tier for quantized 32B models, longer contexts, and local coding or RAG workflows. The operating system and runtime consume part of the shared pool.
Large-model platform guide
Ryzen AI Max+ 395 PCs for local LLMs
Systems with 64 or 128 GB of unified memory can offer more model capacity than typical consumer GPUs. The crucial questions are how much memory the GPU can actually use and whether your runtime supports the Radeon 8060S.
Short answer
When does this platform make sense?
128 GB: the 70B path
The larger configuration creates enough room for many quantized 70B models. Whether a specific model and context fit depends on quantization, the UMA limit, and KV cache.
Capacity before peak speed
The Radeon 8060S shares memory with the CPU. That enables large models in a compact machine, but it does not replace a high-end discrete GPU when maximum token throughput is the priority.
128 GB of unified memory is not 128 GB of free VRAM. Always reserve memory for the operating system, runtime, context, and other applications.
Complete systems
Two 128 GB configurations for different priorities
Both systems use Ryzen AI Max+ 395 and Radeon 8060S. The deciding differences are storage, ports, physical layout, warranty, and final price because the memory is not user-upgradeable.
Best for: large models in a compact PC
GMKtec EVO-X2
Ryzen AI Max+ 395 with 128 GB LPDDR5X unified memory and a 2 TB SSD.
Good match for
Quantized 70B models, larger local knowledge systems, and users who value model capacity more than maximum discrete-GPU speed.
Check before buying
Select exactly 128 GB / 2 TB, confirm the maximum UMA allocation, and verify the operating-system and runtime path you plan to use. The soldered memory cannot be upgraded later.
- CPU / graphics
- Ryzen AI Max+ 395 / Radeon 8060S
- Memory
- 128 GB LPDDR5X, soldered
- Storage expansion
- 2× M.2 PCIe 4.0; 2 TB included
- Networking
- 2.5GbE, Wi-Fi 7
- Ports
- 2× USB4, HDMI, DisplayPort, SD reader
- Size / power
- 193 × 186 × 77 mm; external 230 W adapter
Best for: professional support and Windows
HP Z2 Mini G1a
Ryzen AI Max+ PRO 395 with 128 GB unified memory, a 1 TB SSD, and a three-year workstation warranty.
Good match for
Quantized 70B models, professional local AI work, and buyers who value Thunderbolt 4, an internal power supply, and on-site support more than maximum included storage.
Check before buying
Confirm BN8E8UA#ABA with 128 GB / 1 TB, the intended UMA allocation, and your Radeon runtime. One terabyte is a limited starting point for a large model library, and the workstation package can carry a substantial price premium.
- CPU / graphics
- Ryzen AI Max+ PRO 395 / Radeon 8060S
- Memory
- 128 GB LPDDR5X-8533, soldered
- Storage expansion
- 2× M.2 2280; 1 TB included
- Networking
- 2.5GbE, Wi-Fi 7
- Ports
- 2× Thunderbolt 4, 2× mini-DP, USB-C/A
- Warranty
- 3 years parts, labor, and on-site repair
* Paid link. We may earn a commission if you buy; your price is unchanged. As an Amazon Associate I earn from qualifying purchases. Recheck configuration, seller, and delivery details on the destination page.
Memory decision
64 or 128 GB?
| Question | 64 GB | 128 GB |
|---|---|---|
| Typical focus | quantized 20B–32B | 32B with ample reserve; many 70B quantizations |
| Long contexts | calculate reserve carefully | more headroom, still model-dependent |
| Multiple local services | possible, but pressure rises quickly | better for RAG, embeddings, and parallel components |
| Later RAM upgrade | not possible | not possible |
| Recommendation | when 70B is not essential | when large models drive the purchase |
Before ordering
Seven checks that matter
RAM configuration
LPDDR5X memory is soldered. A 64 GB unit cannot be upgraded to 128 GB later.
UMA/VRAM allocation
Check the BIOS version and maximum GPU allocation. Advertised total memory is not automatically all available to the model.
Runtime and backend
llama.cpp lists HIP and Vulkan GPU backends. AMD's HIP documentation includes Ryzen AI Max+ 395, but your exact runtime release still needs to match.
SSD capacity
Model files, multiple quantizations, and embedding models can consume hundreds of gigabytes. A 2 TB SSD is the more comfortable starting point.
Sustained-load profile
Quiet, balanced, and performance modes use different power limits. Keep the vents clear for long responses or batch jobs; a quieter mode may be preferable in a living or sleeping area.
Ports and networking
Fast Ethernet and enough USB4/display outputs are useful for a NAS, a local model server, or multiple clients.
Returns and warranty
Confirm seller, included power adapter, delivery scope, and support conditions for the exact configuration.
Desk and expansion
Choose storage or workstation support
The EVO-X2 is flatter and includes 2 TB, but uses an external 230 W adapter. The HP Z2 Mini G1a is taller, starts with 1 TB, and adds an internal power supply, Thunderbolt 4, and a three-year workstation warranty. Both provide two M.2 slots, 2.5GbE, Wi-Fi 7, and soldered memory.
Next comparison
Capacity or CUDA speed?
If 32 GB of VRAM is enough for your models, an RTX 5090 tower may be the better direction.
Open RTX 5090 guide