Large models
128 GB of LPDDR5X provides room for many quantized 70B models with usable headroom. The amount assignable to graphics depends on BIOS, operating system, and runtime.
128 GB mini PC with high-speed networking
Beelink GTR9 Pro for local LLMs
The GTR9 Pro combines Ryzen AI Max+ 395 and 128 GB of unified memory with two 10GbE ports. It is especially compelling as a compact model server that delivers large quantized models to several devices on a home network.
Bottom line
When does the GTR9 Pro fit?
Home model server
Two 10GbE ports distinguish the GTR9 Pro from many compact 395 systems. They are useful for fast network storage, multiple clients, or separated data and management networks.
Sustained compact performance
A vapor chamber and dual fans are designed for sustained power. Placement, acoustics, and unobstructed airflow still need to suit your intended workload.
Specific complete system
128 GB / 2 TB configuration
This is the capacity-first version to choose for quantized 70B models, large RAG stores, and networked serving. Its soldered memory cannot be expanded later.
128 GB / 2 TB
Beelink GTR9 Pro
Ryzen AI Max+ 395, Radeon 8060S, 128 GB LPDDR5X-8000, a 2 TB SSD, dual 10GbE, Wi-Fi 7, two USB4 ports, and an SD 4.0 reader.
View 128 GB / 2 TB configuration*Best use
Capacity-focused local server
Good for: quantized 70B models, large RAG collections, a central LLM service on a home network, and buyers who value networking more than the flattest enclosure.
Less suitable for: maximum token speed with small models, silent operation, or software that strictly requires CUDA.
* Paid link. We may earn a commission if you buy; your price is unchanged. As an Amazon Associate we earn from qualifying purchases. Confirm the 128 GB / 2 TB selection, seller, and delivery details before buying.
Hardware at a glance
What matters for local AI
| Component | Configuration | Practical impact |
|---|---|---|
| Processor / graphics | Ryzen AI Max+ 395 / Radeon 8060S | large shared memory pool; Radeon runtime required |
| Memory | 128 GB LPDDR5X-8000 | many quantized 70B models fit; no later upgrade |
| Storage | 2 TB NVMe, two M.2 2280 slots | room for a separate model drive |
| Networking | 2× 10GbE plus Wi-Fi 7 | well suited to NAS and multi-client serving |
| External I/O | 2× USB4, HDMI 2.1, DisplayPort 2.1, SD 4.0 | strong desktop and fast-storage connectivity |
Before buying
Five checks that matter
Exact configuration
The listing must explicitly state 128 GB and 2 TB. Soldered LPDDR5X cannot be replaced later.
Ethernet on Linux
Verify driver support for the 10GbE controller on your intended distribution. Test link stability and power-saving states before unattended serving.
Radeon runtime
Vulkan and HIP/ROCm continue to evolve. Confirm that your application, quantization, and operating-system version support the intended GPU path.
Memory reserve
The operating system, runtime, KV cache, and parallel services consume part of the 128 GB pool. Model weights do not get the full headline capacity.
Placement
Keep airflow paths clear during long inference jobs. In a living space, test power profiles and acoustics under a representative sustained load.
Compare before choosing
GTR9 Pro, EVO-X2, or Z2 Mini?
Choose the GTR9 Pro for dual 10GbE, the EVO-X2 for a flatter capacity-first system, or the HP Z2 Mini G1a when workstation support and an internal power supply matter more.