If the requirement is fitting a 120B-parameter model locally, the GMKtec EVO-X2 is the 128 GB-class mini PC to buy. PCWorld praised its 'excellent combination of CPU, GPU, and NPU performance at desktop workstation level' and scored it 4.5/5. The XDNA 2 NPU contributes 50 TOPS toward a 126 TOPS platform total once the Ryzen AI Max+ 395 and the Radeon 8060S iGPU are counted. With 128 GB of LPDDR5X at 256 GB/s it loads GPT-OSS 120B Q4 with room to spare and runs 70B-class models at 6-8 tokens/sec for a single user. Memory is soldered, and a single 2.5G Ethernet port limits clustering.
Full review
128 GB of LPDDR5X at 256 GB/s
The whole case for this box is memory. 128 GB of LPDDR5X-8000 across a 256-bit bus yields 256 GB/s of bandwidth, which is what lets it load GPT-OSS 120B Q4 at around 70 GB with headroom left over. Single-user Llama-3-70B Q4_K_M decode lands in the 6-8 tokens/sec range, while an 8B model at the same quant runs 35-50.
Ryzen AI Max+ 395 and a 50 TOPS NPU
The processor is AMD's Ryzen AI Max+ 395: 16 cores, 32 threads, a top clock of 5.1 GHz, paired with the 40-core Radeon 8060S iGPU. The XDNA 2 NPU contributes 50 TOPS on its own and 126 TOPS platform-wide once CPU and iGPU are folded in, the same figure Beelink markets on the GTR9 Pro because it is the same Strix Halo silicon.
Desktop workstation power in 7.6 inches
PCWorld's Christoph Hoffmann called it 'a compact workstation with desktop power' rather than an ordinary mini PC, and the 7.6 x 7.3 x 3.03-inch chassis slides behind a monitor or under a riser. Tom's Hardware found performance holds steady under sustained load, and IGN placed the Radeon 8060S in the region of a mobile RTX 4050 to 4060 for gaming.
Removable top panel, but soldered memory
IGN praised the removable top panel for putting the full-size M.2 slot within easy reach, so storage is upgradeable. Memory is not: the LPDDR5X is soldered, unavoidable at this bandwidth but permanent. Dedicated fan-mode buttons on the case are a small piece of control most mini PCs omit, and internal service work beyond the SSD is harder than it should be.
A single 2.5GbE port where the GTR9 Pro has dual 10GbE
Networking is the clearest gap in the specification: one 2.5G Ethernet port. For a single workstation that is fine, but for anyone planning to cluster boxes or pull models off a fast NAS, the Beelink GTR9 Pro's dual 10GbE is the better fit. Wi-Fi 6 and Bluetooth 5.2 round out connectivity, both a generation behind current parts.
Oversized if your model fits in 64 GB
The honest caveat is that 128 GB is more memory than most local-LLM users need. If your working model lives in 70B Q4, a Mac mini M4 Pro is faster per dollar. The EVO-X2 earns its price only when the model genuinely will not fit in 64 GB, and at that point it is the cheapest route to running one on Linux or Windows.
Strengths
- +AMD Ryzen AI Max+ 395 with 50 TOPS XDNA 2 NPU (126 TOPS platform total CPU+GPU+NPU)
- +Supports multiple open-source AI models and popular development frameworks (Ollama, llama.cpp, MLC)
- +Roughly half the price of an Nvidia DGX Spark for similar 128 GB-class local-LLM work
Watch-outs
- −Memory is soldered — no future RAM upgrades
- −Single 2.5G Ethernet port limits AI clustering compared to the Beelink GTR9 Pro
How it compares
The GMKtec EVO-X2 has 128 GB of unified memory at 256 GB/s, which fits 120B Q4 models that the Mac mini M4 Pro 64 GB cannot. The Apple Mac Studio M4 Max reaches the same 128 GB at 546 GB/s, more than double the bandwidth. The Framework Desktop (Ryzen AI Max+ 395) uses the same Ryzen AI Max+ 395 chip with up to 128 GB at 256 GB/s.
Rating sources
Counts as 4.64/5 in our score — this publisher's 60 scores on the site average 4.02, so its scores are adjusted up to compare fairly with other publishers.
Counts as 4.18/5 in our score — this publisher's 95 scores on the site average 3.98, so its scores are adjusted up to compare fairly with other publishers.
Counts as 4.03/5 in our score — this publisher's 15 scores on the site average 4.13, so its scores are adjusted up to compare fairly with other publishers.
Our 4.3 score is the average of the 3 scored reviews above, each adjusted for how strictly its publisher scores; 1 more reviewed it without printing a score and is not in the average. More about methodology.
