GMKtec EVO-X2 AI Mini PC Ryzen Al Max+ 395 Superchip 128GB LPDDR5X 2TB SSD
Más de 100 compradas el mes pasado
US$3,649.99
Precio en USD + envío por peso real. Llega en 5 a 10 días hábiles con Envi Express. Pagá en línea con tarjeta.
Más de 100 compradas el mes pasado
US$3,649.99
Precio en USD + envío por peso real. Llega en 5 a 10 días hábiles con Envi Express. Pagá en línea con tarjeta.
4,1
82 calificaciones
Las calificaciones provienen de Amazon.com.
Amazon Customer
8 de mayo de 2026
Solid AI workstation for running local models
Very fast and performant AI device, great value for money (I bought it on sale when it released last november, current prices are high but then again everything is expensive), I would have liked one more HDMI port so far but aside from that the ports are pretty good connectivity-wise. I found it to be quite portable and brought it around on my travels with a portable monitor as well. I think the fans could be better, not very noisy but could be more efficient in cooling down the machine, but really what you are paying for is the power of the strix halo.Regarding LLM models, it can run most models that are sub 70GB so 120-130B parameter models run at Q4_KS or IQ4_XS... if they are MoE I can get about 12 t/s for inference and dense can be quite slow. 70B~ is around 3 t/s, haven't tried the latest mistral dense yet. Anything smaller than those sizes is blazingly fast for inference, so I can run some webapps and overhead as well as the models. I'd highly recommend it for any LLM hobbyist.
lance b
A token generating beast!
This system was spendy but worth every penny. I have been running it 24/7 since I got it. Right now I have local AI LLM accessible from any pc or phone on my local network. I have Hermes agent harness working on multiple projects at a time and soon with the VPN I'm setting up I will be able to access my GMKTec AI LLM server from anywhere I have internet access. This computer is a token generating beast. I was worried about it getting to hot so I bought a small desktop fan and have it blowing on it. But after running it for awhile now my concerns about heat ended up not being a problem. It gets warm, but barely ever gets what I would consider hot. I have ran models up to 70b in size so far. The larger the model the slower it will be able to process tokens, but it can handle quite large LLM models. My favorite go to model to use on this machine is Orinth 35b with Q4KM quantization. I did buy a surge protector UPS for it to protect from voltage dips, spikes, and power loss since this was a big investment for me. This computer is a dream come true!
Mike
19 de marzo de 2026
Great value for this form factor - work required if really want to use it for AI
I bought the 128 GB configuration to run large language models locally. After a month of daily use as a dedicated AI inference server on Linux, this is a very capable mini PC, but it's not plug-and-play for this use case.THE SHORT(ISH) VERSIONThe 128 GB unified memory is the killer feature. For local AI inference, memory is everything. The Ryzen AI Max+ 395 with its integrated Radeon 8060S (40 compute units, RDNA 3.5) is fast enough to make running serious models practical, not just possible.What I love:- 128 GB unified memory lets me run models that would require a high-end discrete GPU elsewhere- Dual NVMe slots, it came with a 2 TB drive and I added a Crucial T710 4 TB NVMe for model storage. 6 TB total, no external drives needed. Note: the slots are PCIe 4.0, so Gen5 drives like the T710 work but run at Gen4 speeds- Cooling hasn't been an issue, I run it 24/7 under loads with no problems- I run it headless as a dedicated server, 2.5 GbE and SSH is all I need. Ports haven't been a concern for my use case- The footprint is tiny. It sits on my desk under my monitorsWhat to know before buying:- RAM is soldered. Get the 128 GB config. There is no upgrade path.- If you're buying this for AI/ML on Linux, read the technical section below- The marketing focuses on Windows AI features and the NPU, the real value for AI workloads is the GPU + unified memory on LinuxSTAR RATING: 4/5. The hardware is genuinely exceptional for the price. I dock one star because the AI use case, which is the main reason to buy the 128 GB config, requires significant Linux configuration that isn't documented by GMKtec. They're selling AI-capable hardware with consumer-grade setup guidance.FOR AI/ML USERS ON LINUXIf you're here because you want to run local LLMs, here's what I've learned running this machine 24/7 for a month. GMKtec's listing mentions LM Studio and Deepseek 70B, my results go well beyond that.WHAT IT CAN ACTUALLY DOReal benchmarks from my setup running llama.cpp with ROCm 7.2.0 on Ubuntu:- Qwen3.5-35B (3.5B active MoE, Q6_K, 28 GB): 44 tokens/sec generation, 180-300 tok/s prompt processing via Vulkan- Qwen3-Coder-Next 80B (MoE, Q4_K_M, 46 GB): 42 tokens/sec, an 80B coding model at conversational speed- Qwen3.5-122B (10B active MoE, Q4_K_M, 75 GB): 15 tokens/sec with HIP backend, a 122-billion parameter model running on a mini PC- I run whisper.cpp for speech-to-text (~1.5 GB) alongside the LLM server simultaneously with no issues- 8 parallel inference slots with 262K context window on the 35B modelThese aren't cherry-picked numbers. This is what I get daily serving models to multiple clients on my network.WHAT REQUIRED WORKThis is not plug-and-play for AI workloads. You need GRUB kernel parameters for ROCm stability:- amdgpu.cwsr_enable=0 (prevents GPU crashes)- amd_iommu=off (IOMMU compatibility)- amdgpu.gttsize=126976 (expands the GPU translation table to 124 GB, the stock default is much smaller)- ttm.pages_limit=32505856 (memory management)Without these, expect crashes and out-of-memory errors that don't make sense given the hardware specs. I spent real time figuring this out.GTT fragmentation is real, even with 124 GB configured, contiguous allocation limits mean you can't always use it all. The 122B model requires careful batch-size tuning (I dropped from 16384 to 2048) to fit. Full GPU offload (ngl=99) still OOMs on the 122B; I run it at ngl=44.Vulkan vs. HIP matters: Vulkan is ~10% faster for MoE models (the sparse ones), HIP is ~10% faster for the 122B dense-ish model. You'll want to benchmark both for your specific models.WHO SHOULD BUY THISIf you want to run 30-80B parameter models at usable speeds without a discrete GPU, this is a really good option in this form factor. The unified memory architecture means the GPU can access the full 128 GB, something that would require a high-end discrete GPU or multiple consumer GPUs to match.If you just want a fast mini PC for normal desktop use, you're paying a premium for 128 GB of soldered RAM you may never use. Consider the 96 GB config instead.If you want a turnkey AI appliance, this isn't it. Budget time for Linux setup and GRUB tuning. But if you're comfortable with that, this machine punches absurdly above its weight class.COMMUNITY RESOURCESGMKtec doesn't document the Linux AI setup, but the community does. These were invaluable for getting this machine dialed in:- kyuz0/amd-strix-halo-toolboxes (GitHub) GRUB parameters, flash attention, no-mmap recommendations, firmware warnings- pablo-ross/strix-halo-gmktec-evo-x2 (GitHub) EVO-X2 specific Ubuntu setup guide- strixhalo.wiki community performance tracking and tuning tips for Strix HaloStart there. It'll save you hours.
James Kimber
4 de abril de 2026
Blazing fast AI workstation that works perfect with Fedora
Been running for about a month with no issues. Fantastic AI workstation and quite a bit heavier than I expected! Booted into Windows very quickly, then after verifying BIOS was up to date, wiped Windows and installed the latest beta of Fedora 44. WiFi, Ethernet, Bluetooth immediately recognized and connected. Ethernet on Realtek RTL8125 and WiFi uses Mediatek MT7925, so double check compatibility if you're installing Linux. Screaming fast and cleanly running 90b local AI models with Open WebUI and SearXNG. If you're pushing this box to the limit, keep it well ventilated and in a cool place because it can get hot quickly when running heavier AI operations. Otherwise it's near silent and CPU stays around 36°C for most operations with a really low power draw