GMKtec EVO-X3 Review: The Strix Halo Mini PC With OCuLink

What Is the GMKtec EVO-X3?

The GMKtec EVO-X3 is the successor to the EVO-X2, rebuilt with a completely new metal chassis, triple-fan cooling, a 140W performance mode and one feature no other Strix Halo mini PC has: an OCuLink port for connecting an external desktop GPU.

GMKtec launched the EVO-X2 in early 2025 at around $1,999, and it quickly became a reference point for Strix Halo mini PCs. The EVO-X3 arrived in mid-2026 with the same Ryzen AI Max+ 395 APU — still the flagship Strix Halo chip — but wrapped in a substantially different machine. The old plastic-and-aluminum sandwich design is gone, replaced by a taller, tower-style metal chassis that Tom’s Hardware describes as “almost entirely metal” with a CNC-machined body and a textured green top cap.

The redesign is not purely cosmetic. The extra height accommodates larger fans, better exhaust pathways and more thermal headroom, which lets GMKtec push the APU to a 140W peak power mode — up from the EVO-X2’s more conservative limits. And the OCuLink port, which was absent on the EVO-X2, transforms the EVO-X3 from a fixed-configuration mini PC into a platform you can expand with a discrete GPU later.

What hasn’t changed: 128GB of onboard LPDDR5X-8000 unified memory, up to 96GB of which can be allocated as VRAM to the Radeon 8060S GPU. That’s the configuration that matters for local AI, and it’s identical to the BOSGAME M5 and the DGX Spark 128GB.

Full Specifications

The OCuLink Advantage: What Makes the EVO-X3 Different

The EVO-X3 is the only Strix Halo mini PC with a built-in OCuLink port. That PCIe Gen4 x4 connection delivers far more bandwidth than USB4 for an external GPU — enough for Wendell of Level1Techs to plug an NVIDIA GPU and an AMD GPU into the same Strix Halo machine and run both simultaneously.

This is the EVO-X3’s defining feature. Every other mini PC in this class — the BOSGAME M5, the Minisforum MS-S1 Max, the Beelink GTR9 Pro — relies on USB4 or Thunderbolt for external expansion, which caps out at around 40 Gbps. OCuLink over PCIe Gen4 x4 offers 64 Gbps of raw bandwidth, and more importantly, it exposes the GPU over a PCIe link rather than a tunneled protocol. In practice, that means lower latency and better performance for eGPU workloads.

Level1Techs demonstrated something more interesting than a simple eGPU test. By combining the EVO-X3’s internal Radeon 8060S with an external NVIDIA card over OCuLink, they ran a hybrid configuration where the AMD GPU handled one part of the workload and the NVIDIA GPU handled another — in their case, generating about 87 tokens per second on the NVIDIA card while the AMD chip contributed about 24 tokens per second in parallel. The local model was even used to debug the cluster itself.

There’s a caveat worth stating clearly: PCIe Gen4 x4 will throttle modern high-end GPUs. An RTX 5090 running over OCuLink won’t deliver the same performance as it would in a full x16 slot. But for the kind of workloads this machine targets — accelerating prefill, adding CUDA capability for specific tools, or offloading a model that doesn’t fit in the 96GB VRAM allocation — OCuLink is a genuine and unique capability.

Local AI Performance: ~60 Tokens/Second on Qwen 3.6

With 128GB of LPDDR5X-8000 and up to 96GB allocatable to the GPU, the EVO-X3 runs 120B-class models locally. Level1Techs measured Qwen 3.6 at nearly 60 tokens per second on Windows using AMD’s Lemonade stack. A user report on GMKtec’s Japanese store shows Qwen 3.6 35B running at about 45 tokens per second in llama.cpp on Ubuntu.

The 140W performance mode deserves context. Wendell made the most useful observation in his review: for token generation, the CPU power limit doesn’t matter much — what matters is memory speed. The 140W mode helps with prefill and prompt processing, where compute matters more. For generation, you’re bound by the 8000 MT/s memory and the 256 GB/s memory bandwidth, not by how many watts the CPU is pulling. His AIDA64 run measured 118 GB/s of real-world memory bandwidth, which he compared to a quad-channel Threadripper system — impressive for a mini PC.

In his benchmark suite, Wendell also noted that the EVO-X3’s 16 Zen 5 cores “run circles around the 20 cores you get in the DGX Spark for CPU-type operations.” For mixed workloads that blend inference with traditional CPU compute, that’s a meaningful advantage.

• Qwen 3.6 (large model): ~60 tok/s on Windows with Lemonade stack (Level1Techs measurement)
• Qwen 3.6 35B: ~45 tok/s in llama.cpp on Ubuntu (user report, GMKtec JP store)
• 8B models: ~40 tok/s in hybrid mode (Level1Techs forum data)
• Time to first token: ~0.99 seconds on optimized setups, though an 18,000-token prompt took about 44 seconds to reach the first answer on a tuned 128GB Strix Halo configuration

The software situation is straightforward. Windows works, and AMD’s Lemonade stack provides a one-install path to the GPU and NPU. But Linux remains the better path for local AI — more control over memory allocation, and the BIOS now exposes SR-IOV, which lets you partition GPU resources for virtualization. GMKtec also ships its own Claw + Wrangler software for local AI agents, though most users in the local AI community will stick with Ollama or LM Studio.

Gaming and SteamOS: A Surprisingly Capable Machine

The Radeon 8060S is AMD’s fastest integrated GPU, and the EVO-X3’s 140W power budget lets it stretch further than most mini PCs. ETA Prime tested the machine with SteamOS and measured Cyberpunk 2077 at nearly 80 FPS at 1440p with high settings and FSR Quality. Horizon Zero Dawn Remastered and Spider-Man 2 both averaged over 80–85 FPS at 1440p.

SteamOS and the EVO-X3 are, in the words of Notebookcheck’s coverage, “a perfect combination.” Linux-based operating systems often extract better performance from AMD hardware than Windows does, and the EVO-X3’s triple-fan cooling keeps thermals under control during extended gaming sessions. Tom’s Hardware noted the machine stays “quiet under load” — the fans ramp up but remain unobtrusive, even at the 140W performance mode.

The gaming capability matters for a different reason: it makes the EVO-X3 a genuine all-in-one machine. You can run local LLM inference during the day, switch to a game in the evening, and do creative work in between — all on the same box. The DGX Spark can’t do that; it’s a Linux-only AI appliance with no gaming capability.

GMKtec EVO-X3 vs BOSGAME M5 vs DGX Spark 128GB

The EVO-X3 occupies a distinct position. It’s $800 more than the BOSGAME M5 but adds OCuLink, a larger metal chassis, a 140W performance mode and triple-fan cooling. Against the DGX Spark 128GB, it’s $3,150 cheaper and far more versatile as a general-purpose machine — but it lacks CUDA, FP4 acceleration and ConnectX-7 clustering.

If you’re building a local AI setup and trying to decide between Strix Halo and the Spark, our full comparison of the DGX Spark 64GB, 128GB and Strix Halo covers the decision tree in detail. And if you’re new to the platform, our Strix Halo explainer covers the architecture and what makes it suitable for local AI.

Verdict: Should You Buy the GMKtec EVO-X3?

Buy the EVO-X3 if you want OCuLink, a premium metal chassis, and the flexibility to add an external GPU later. Buy the BOSGAME M5 instead if you want the cheapest path to 128GB of unified memory and don’t need eGPU expansion. Skip both if your workflow depends on CUDA or you need NVIDIA’s clustering capability.

Buy the GMKtec EVO-X3 ($3,799) if…

• You want the OCuLink port — it’s the only Strix Halo mini PC with one, and it opens the door to external GPU acceleration later
• You want a premium metal chassis with triple-fan cooling and a quiet experience under heavy AI loads
• You want to run SteamOS for gaming alongside your AI workloads — the EVO-X3 is exceptionally well-suited to it
• You value build quality and thermals enough to pay $800 more than the M5

Buy the BOSGAME M5 ($2,999) instead if…

• Price per GB is your priority — the M5 is the cheapest 128GB Strix Halo mini PC on the market
• You want a built-in SD card reader — the M5 has one, the EVO-X3 doesn’t
• You don’t need OCuLink and prefer a smaller, lighter machine

Consider the DGX Spark 128GB ($6,950) if…

• Your workflow is CUDA-native — professional ML development or NVIDIA AI Enterprise
• You need FP4 acceleration or ConnectX-7 clustering
• Budget is not the primary constraint

Frequently Asked Questions