GMKtec EVO-X2 (Ryzen AI Max+ 395 128GB)

The GMKtec EVO-X2 (Ryzen AI Max+ 395 128GB) is a flagship mini PC purpose-built for local AI inference. It leverages AMD’s Strix Halo APU, a 16-core Zen 5 CPU paired with a Radeon 8060S iGPU and XDNA 2 NPU, to deliver 96 GB of unified GPU-allocatable memory in a compact, low-power chassis. At $1,999, it occupies the prosumer tier: more capable than a consumer laptop, far more energy-efficient than a multi-GPU workstation, and directly competitive with entry-level server-grade inference rigs.

This is not a general-purpose desktop. The EVO-X2 is engineered for one task: running large language models and other AI workloads locally, without cloud dependency. Its 128 GB of LPDDR5X-8000 unified memory can allocate up to 96 GB to the GPU, enabling models that would otherwise require a $10,000+ setup. The design won the American Good Design Platinum Award 2025, reflecting its industrial and thermal engineering, a metal chassis with triple-fan cooling keeps sustained loads under 70 °C.

For developers and researchers who need to run 70B-parameter models at Q5 quantization, or experiment with sparse 235B models, the EVO-X2 is currently the most accessible hardware that can do so in a desk-friendly form factor.

The headline spec is the 96 GB of GPU-accessible memory. This is not a discrete GPU with dedicated VRAM; it’s unified memory shared between the CPU and GPU via the Radeon 8060S iGPU (40 RDNA 3.5 compute units). The full 96 GB allocation is unlocked by default on this SKU, no BIOS tweaks or memory reservation tricks required. This is the largest unified memory pool available in any mini PC as of mid-2025.

The 128 GB of LPDDR5X-8000 runs on a 256-bit bus, delivering 256 GB/s bandwidth. That’s roughly 60% of a desktop RTX 4090 (1 TB/s) but more than enough for token generation at reasonable batch sizes. For context, a Mac M2 Ultra with 192 GB unified memory peaks at 800 GB/s, but costs over $5,000. The EVO-X2’s bandwidth is a limiting factor for very large models (235B+), but for the 70B–120B range it’s well-matched to the compute throughput.

The XDNA 2 NPU provides 50 TOPS for integer operations, but the real workhorse for LLM inference is the iGPU’s FP16/INT4 matrix units. The Radeon 8060S iGPU delivers approximately 12.5 TFLOPS (FP16), comparable to an RTX 4060 Ti. Combined with the CPU’s 16 Zen 5 cores (boost up to 5.1 GHz), the system achieves ~50 tokens/s on 7B models and ~23 tokens/s on 20B models in real-world benchmarks (e.g., GPT-OSS-20B at 23.8 tok/s out of the box on Ubuntu).

At 140W peak TDP, the EVO-X2 consumes less than a third of a typical gaming GPU. This makes it viable for edge deployment, home labs running 24/7, or offices where noise and heat matter. The triple-fan cooling keeps the APU under 70 °C during sustained inference, though brief throttling may occur after 15 minutes of peak load (per Mini PC Reviewer tests).

• APU: AMD Ryzen AI Max+ 395 (16C/32T Zen 5, Radeon 8060S 40 CUs, XDNA 2 NPU)
• Memory: 128 GB LPDDR5X-8000 (256-bit, 256 GB/s)
• GPU VRAM Allocation: Up to 96 GB
• Storage: 2 TB PCIe 4.0 NVMe (Crucial P3)
• Connectivity: USB4, Wi-Fi 7, BT 5.4, SD Card 4.0, dual LAN (2.5 GbE)
• Display: Quad 8K via HDMI 2.1 / DisplayPort 1.4
• OS: Windows 11 Pro (also tested with Ubuntu 24.04.3 LTS)

This is the critical question for any practitioner. The EVO-X2’s 96 GB unified memory and 256 GB/s bandwidth define a clear capability envelope.

• 70B at Q5_K_M: This is the ideal workload. Models like Llama 3.1-70B, Qwen 2.5-72B, or Mistral Large 2 fit comfortably with ~48 GB VRAM usage, leaving room for context (up to 128K tokens). Expect 10–15 tok/s on a single prompt.
• 120B at Q4: Models such as DeepSeek-R1-120B or Mixtral 8x22B (MoE) can be run with careful context management. Real-world tests show ~15 tok/s for GPT-OSS-120B.
• 235B Sparse: Qwen3-235B (sparse MoE) can be loaded at ~50% density, using nearly all 96 GB. Token generation slows to ~11 tok/s, usable for batch inference or offline analysis, but not real-time chat.

The unified memory architecture excels at multimodal models (e.g., LLaVA, Qwen-VL) because image embeddings share the same pool without PCIe transfers. Long-context tasks (128K+ tokens) are feasible for models up to 70B; beyond that, context length must be reduced to avoid OOM.

Based on benchmarks from nishtahir.com and Mini PC Reviewer:

• 7B Q4_K_M: ~50–55 tok/s (prompt eval >1900 tok/s)
• 20B Q5_K_M: ~23 tok/s
• 70B Q5_K_M: ~11–14 tok/s
• 120B Q4: ~5–8 tok/s (prompt eval slower due to bandwidth)

For production inference servers, batch size 1 is typical; for hobbyist use, these speeds are comfortable for interactive chat on models up to 70B.

• Local LLM Enthusiasts who want to run 70B models without cloud subscriptions or noisy GPU rigs. The EVO-X2 is a silent, compact alternative to a dual RTX 4090 setup.
• AI Application Developers building agentic workflows that require private, low-latency inference. The 96 GB unified memory allows serving multiple models simultaneously (e.g., a 7B router + 70B generator).
• Edge AI Deployments where power and space are constrained but model size cannot be compromised. At 140W peak, it can run off a standard outlet in a retail kiosk, medical device, or field lab.
• Research Teams needing to fine-tune or experiment with 70B–120B models on a single machine. While not designed for training (no discrete GPU), the EVO-X2 can handle LoRA fine-tuning on small batches.

• High-throughput inference servers serving hundreds of concurrent users, you need multiple GPUs with higher bandwidth.
• Training large models from scratchthe iGPU’s 12.5 TFLOPS is insufficient for pre-training.
• Gamersthe Radeon 8060S can handle 1440p at medium settings, but a discrete GPU is better for high-refresh gaming.

• Price: EVO-X2 $1,999 vs. Mac Studio ~$5,000+ for 192 GB.
• VRAM: 96 GB vs. 192 GB, Mac Studio wins for 120B+ models.
• Bandwidth: 256 GB/s vs. 800 GB/s, Mac Studio is faster for token generation.
• Software: EVO-X2 runs Windows/Linux with full CUDA/ROCm compatibility via Vulkan/LLVM; Mac Studio limited to Metal and Apple’s stack.
• Pick EVO-X2 if: you need Linux-native tooling (llama.cpp, vLLM, Ollama) and a lower budget.

• VRAM: 96 GB vs. 24 GB, EVO-X2 can run 70B models that a 4090 cannot fit.
• Speed: 4090 generates tokens ~2–3× faster for models that fit (e.g., 7B at 150 tok/s).
• Power: 140W vs. 450W, EVO-X2 is far more efficient.
• Pick EVO-X2 if: your models exceed 24 GB VRAM, or you need a silent, compact system.

• VRAM: 96 GB vs. 64 GB, EVO-X2 supports larger models.
• Bandwidth: 256 GB/s vs. 273 GB/s, comparable.
• Price: $1,999 vs. $1,399, Mac Mini is cheaper for smaller models.
• Pick EVO-X2 if: you need to run 70B+ models or prefer x86 ecosystem.

The GMKtec EVO-X2 is currently the most cost-effective way to get 96 GB of unified memory for AI inference in a compact, low-power package. It fills a gap between consumer laptops and enterprise workstations, making large local models accessible to individual developers and small teams.