EVO-X3

The EVO-X3 is GMKtec's second-generation Strix Halo mini workstation, built around AMD's Ryzen AI Max+ 395. It replaces the EVO-X2 with a redesigned full-metal CNC chassis, a beefier cooling system, and 128GB of LPDDR5X-8000 memory as standard. At $3,799.99 with a 2TB or 4TB NVMe drive, it sits firmly in the prosumer tier — the same bracket as Apple's Mac Studio M3 Ultra and high-end discrete-GPU workstations like a dual RTX 4090 build, but in a 353 x 186 x 41 mm, 2.3 kg enclosure.

What makes the EVO-X3 interesting for AI work isn't raw compute — it's the memory architecture. The Ryzen AI Max+ 395 uses a unified memory pool, so up to 96GB of the 128GB LPDDR5X can be carved out as VRAM for the Radeon 8060S iGPU. That's the headline number: 96GB of GPU-addressable memory in a mini PC, which is enough to load quantized 200B+ class models that simply won't fit on a single 24GB or 48GB consumer GPU. GMKtec lists support for LLMs up to 235B parameters.

For practitioners weighing the EVO-X3 for AI, the tradeoff is clear. You get enormous memory capacity and quiet operation at 140W peak, but you pay a premium and accept integrated-graphics-tier compute throughput. This is a capacity play, not a speed play.

The specs that matter for local inference:

• VRAM: 96GB allocatable from 128GB LPDDR5X-8000 unified memory (non-upgradeable, soldered)
• Compute: 126 TOPS total — 50 TOPS from the XDNA 2 NPU, remainder from the Radeon 8060S (40 RDNA 3.5 CUs, up to 2900 MHz)
• Memory bandwidth: LPDDR5X-8000 — this is the single biggest determinant of token generation speed on a unified-memory system
• TDP: 140W peak, with 85W balanced and 54W silent modes
• CPU: 16 Zen 5 cores / 32 threads, up to 5.1 GHz boost
• Expansion: OCuLink PCIe Gen4 x4 for an external GPU; dual M.2 2280 slots up to 16TB

The 50 TOPS NPU is useful for offloading specific workloads — Windows Studio effects, small on-device models via ONNX Runtime or Ryzen AI toolchains — but most LLM inference on Strix Halo runs on the Radeon 8060S through ROCm or Vulkan backends. The NPU isn't the LLM workhorse here; the iGPU and unified memory are.

Memory bandwidth is the constraint to understand. LPDDR5X-8000 delivers roughly 256 GB/s of theoretical bandwidth shared between CPU and GPU. That's excellent for an integrated part and comparable to an M4 Pro, but it's a fraction of what a discrete RTX 4090 (1 TB/s) or an H100 (3.35 TB/s) provides. The practical consequence: prompt processing (prefill) is fast, but token generation (decode) will be bandwidth-limited. Expect the EVO-X3 to be competitive on time-to-first-token and slower on sustained tokens/sec than discrete-GPU alternatives.

The 140W peak mode is worth enabling for sustained inference. The chassis handles it quietly — Tom's Hardware and TechRadar both flagged the cooling as a strength, with the machine staying quiet under load.

This is where the EVO-X3 justifies its price. With 96GB of VRAM, the model compatibility envelope is far wider than any single consumer GPU.

Comfortably fits (fast, high quality):

• Llama 3.1 70B at Q8 — ~70GB, runs entirely in VRAM with headroom
• Qwen 2.5 72B at Q8, or 32B at FP16
• Mixtral 8x22B at Q6/Q8
• DeepSeek-R1-Distill-Llama-70B at Q8
• Mistral Large 123B at Q4/Q5
• Command R+ 104B at Q5

Fits with careful quantization:

• Llama 3.1 405B at Q2 — technically fits but quality degrades noticeably at 2-bit
• DeepSeek-V3 / R1 (671B MoE) — only with aggressive quantization and CPU offload; expect slow decode
• 235B-class MoE models (Qwen3-235B, etc.) at Q4 — this is GMKtec's stated ceiling and roughly the realistic upper bound for usable quality

Sweet spot: For most practitioners, Q4_K_M or Q5_K_M quantization of 70B–123B models offers the best quality-to-speed tradeoff. A 70B model at Q4_K_M (~40GB) leaves plenty of VRAM headroom for long context windows — 32K to 128K tokens is realistic without paging.

Multimodal and long-context: The unified memory pool makes the EVO-X3 unusually capable for vision-language models (Qwen2-VL, Llama 3.2 Vision, InternVL) and long-context work. Loading a 70B model plus a large KV cache for 128K context is feasible in a way it isn't on a 24GB card.

Expected tokens/sec: Community benchmarks on Strix Halo with 128GB typically land in the 5–10 tok/s range for 70B-class models at Q4, and 15–25 tok/s for 30B-class models. Smaller 7B–14B models run at 40+ tok/s. These are estimates — actual throughput depends on your runtime (llama.cpp, Ollama, LM Studio, vLLM), quantization, and context length.

Developers building agentic workflows. The 96GB VRAM pool lets you run a capable 70B model as a planner alongside smaller specialized models, all resident in memory simultaneously. For "best hardware for local AI agents 2026" searches, the EVO-X3's capacity is its differentiator — you're not juggling model loads.

Researchers and hobbyists running large local LLMs. If you want to run DeepSeek-R1-70B or Qwen 2.5 72B at high quality without a multi-GPU rig, this is one of the few single-box options under $4K.

Teams running inference servers. The 140W ceiling and quiet operation make it viable in an office or lab. OCuLink gives you a path to add a discrete GPU later if you need more throughput.

Edge deployment. Full metal chassis, 2.3 kg, Windows 11 Pro or Ubuntu. It's portable enough for field deployments where a rack isn't available.

Not for training. The 256 GB/s bandwidth and integrated GPU make this an inference machine. Fine-tuning a 7B model with LoRA is possible; anything larger is impractical.

vs. Apple Mac Studio M3 Ultra (128GB/256GB): The Mac Studio offers higher memory bandwidth (up to 819 GB/s on M3 Ultra) and a more mature software stack for MLX-based inference. It typically delivers faster token generation. The EVO-X3 counters with x86 compatibility, CUDA-alternative ROCm tooling, OCuLink expansion, and native Windows/Linux support. If your stack is MLX or you prioritize tok/s, the Mac wins. If you need x86 or plan to add an eGPU, the EVO-X3 is the pick.

vs. dual RTX 4090 build (~$4,500+): Two 4090s give you 48GB VRAM and dramatically higher compute throughput — 5–10x faster token generation on models that fit. But 48GB caps you at ~70B Q4, and you're dealing with 900W+ of heat and noise. The EVO-X3 runs 128B+ models in a 2.3 kg box at 140W. Different tools for different jobs: the 4090 rig is for speed on mid-size models, the EVO-X3 is for capacity on large ones.

vs. GMKtec EVO-X2: Same Ryzen AI Max+ 395 silicon, but the X3 adds the new chassis, standard 128GB LPDDR5X-8000, and improved cooling. The X2 launched at roughly half the price — the X3's premium reflects the 2026 component shortage as much as the redesign.

Pick the EVO-X3 if your workload is memory-bound, not compute-bound — large quantized models, long context, multimodal, or multi-model agent stacks. Skip it if you need maximum tokens/sec on models that already fit in 24–48GB.