Local AI / Hardware

Best Strix Halo Mini PC for Local AI

Buy a GMKtec EVO-X2 if you want the most proven Strix Halo box for local AI and you want it in stock. Every mini PC here runs the same chip, the AMD Ryzen AI Max+ 395: 16 Zen 5 cores, a 40-compute-unit Radeon 8060S iGPU, and up to 128GB of LPDDR5X soldered as one pool the CPU and GPU share. That shared pool is the whole point. It lets a small, quiet box hold a 70-billion-parameter model that will not fit on a 16GB or 24GB consumer graphics card, which is why homelabbers are buying these instead of building a GPU rig. The catch is bandwidth and software: memory tops out around 256 GB/s, far below a discrete card, and there is no CUDA, so plan on Ollama, LM Studio, or llama.cpp over ROCm or Vulkan. Pick the box by what you need around the chip: the EVO-X2 for value and availability, the Framework Desktop for Linux and repairability, the Minisforum MS-S1 Max for dual 10GbE and a real expansion slot.

Updated 2026-10-29 · by

This page has affiliate links, so HomelabCompass may earn a commission at no cost to you. As an Amazon Associate, HomelabCompass earns from qualifying purchases. It never changes a verdict. How the site is funded

Top pick

GMKtec EVO-X2

The EVO-X2 is the box most people actually run Strix Halo on, and that track record matters when the platform is this new. It ships the full Ryzen AI Max+ 395 with the 40-CU iGPU, takes the full 128GB of LPDDR5X, has USB4, dual 2.5GbE, and Wi-Fi 7, and it is the unit most of the local-LLM writeups and YouTube benchmarks were done on. GMKtec also sells it with 32GB, 64GB, or 128GB, so you can match the memory to the models you actually plan to load rather than paying for a pool you will not fill. For local AI it holds its own against every box here, it is widely in stock on Amazon, and it usually undercuts the workstation-style units. You give up a PCIe slot and 10-gigabit networking, and the soldered memory means you choose capacity once, but for a dedicated inference box that is the right trade.

The options

GMKtec EVO-X2

New, roughly $1,800 (64GB) to $3,500 (128GB/2TB), moved up by the 2025 to 2026 memory price surge

The reference Strix Halo mini PC for local AI: a Ryzen AI Max+ 395 box in 32GB, 64GB, or 128GB trims, and the unit most LLM benchmarks were run on.

Visit GMKtec EVO-X2 site
Runs on
Ryzen AI Max+ 395 (16 Zen 5 cores, Radeon 8060S 40 CU), up to 128GB LPDDR5X-8000, dual NVMe, USB4, dual 2.5GbE, Wi-Fi 7
Pick it if
You want the proven, in-stock choice for running local models and the best value in the category. The default pick.

Framework Desktop

New, about $1,100 barebones at the low memory tier up to roughly $2,500 for a 128GB config, storage and OS extra

The same chip in a repairable, Linux-first Mini-ITX box with a standard board you can pull out and reuse later, plus swappable front tiles.

Visit Framework Desktop site
Runs on
Ryzen AI Max+ 395, up to 128GB LPDDR5X-8000 (soldered), 2x M.2 PCIe 4.0 x4, 2.5GbE, Wi-Fi 7, a standard Mini-ITX mainboard and FlexATX PSU
Pick it if
You run Linux, you care about repair and documentation, and you want a board you can redeploy. The pick for openness.

Minisforum MS-S1 Max

New, roughly $2,099 to $2,599 depending on memory and storage

A workstation-style Strix Halo box with a built-in power supply, a full-length PCIe slot, and server networking.

Visit Minisforum MS-S1 Max site
Runs on
Ryzen AI Max+ 395, 128GB LPDDR5X, 2TB Gen4 SSD options, a PCIe x16 physical slot (wired x4), USB4 V2 at 80Gbps, dual 10GbE, internal PSU
Pick it if
You want this box inside a homelab: dual 10-gigabit networking, a slot for a NIC or HBA, and no external power brick.

Bosgame M5

New, often the lowest 128GB price, bundled with a 2TB SSD on most listings

A no-frills 128GB Strix Halo box that tends to be the cheapest way into the full memory tier.

Visit Bosgame M5 site
Runs on
Ryzen AI Max+ 395, 128GB LPDDR5X, 2TB NVMe, USB4, single 2.5GbE, Wi-Fi 7
Pick it if
You want maximum unified memory for the least money and you do not need 10GbE or an expansion slot.

HP Z2 Mini G1a

New, a clear premium over the boutique boxes, often well past $3,000 configured

A first-party workstation on the Ryzen AI Max PRO version of the chip, with business warranty and vendor support.

Visit HP Z2 Mini G1a site
Runs on
Ryzen AI Max+ PRO 395, up to 128GB LPDDR5X, HP's workstation chassis, commercial warranty, manageability and ISV certifications
Pick it if
You need a named vendor, a real warranty, and workstation support, and you will pay for it. The buy-from-a-vendor pick.

How to choose

Choose the memory tier first, because it is soldered and you cannot add more later. The unified pool is the reason to buy one of these, so size it to the models you want to run. A 70-billion-parameter model at a 4-bit quant needs roughly 40GB plus context, which a 64GB box handles with room for the operating system. Step up to 128GB if you want headroom for a large mixture-of-experts model like gpt-oss-120b, a big context window, or a model loaded alongside other services. Do not buy 32GB for AI; at that point a used mini PC and a 16GB graphics card make more sense, and the best mini PC for a home server roundup covers that path.

Be honest about bandwidth and software before you expect GPU speeds. Strix Halo shares one LPDDR5X pool at around 256 GB/s, which is a fraction of what a discrete card pushes, so dense 70B models generate text at a usable but modest pace, roughly mid-single-digit to low-teens tokens per second, not the dozens a 4090 hits on a model small enough to fit. Mixture-of-experts models feel much faster because only a few billion parameters are active per token, which is why gpt-oss and Qwen MoE builds are the sweet spot here. On software, there is no CUDA: you run Ollama or LM Studio on top of ROCm, Vulkan, or llama.cpp. That stack works well now, but it is younger than NVIDIA's, so expect the occasional driver or build note, and favor Linux if you want the smoothest ROCm experience.

Then pick the box by what sits around the chip. For a dedicated inference appliance that lives on a shelf, the GMKtec EVO-X2 or the Bosgame M5 give you the memory and little else to pay for. If the box joins an actual homelab, the Minisforum MS-S1 Max earns its price with dual 10GbE and a PCIe slot for a NIC or HBA, which matters if you are moving models and datasets around a fast network or running it as a Proxmox node on the side. If you run Linux and value repair and a reusable board, the Framework Desktop is the one to get. These boxes idle in the low tens of watts and pull up to about 140 watts flat out, so they are cheap to leave on; put real numbers on it with the power and running-cost calculator.

The verdict

Buy the GMKtec EVO-X2. It is the Strix Halo box most people run local AI on, it takes the full 128GB of unified memory, and it is usually both in stock and cheaper than the workstation units, which is what you want from a dedicated inference appliance. Choose the memory tier carefully, because it is soldered for good: 64GB covers a 70B model, 128GB gives you headroom for a large mixture-of-experts model or a big context window. Get the Framework Desktop instead if you run Linux and want a repairable, reusable board, and the Minisforum MS-S1 Max if the box is joining a homelab and you want dual 10GbE and a real PCIe slot. Go in knowing the tradeoff: around 256 GB/s of bandwidth and no CUDA mean this is a run-big-models-slowly machine, not a speed demon, with Ollama or LM Studio on ROCm or Vulkan doing the work. If you only need a general home server rather than an AI box, you will spend far less on a used micro from the best mini PC for a home server roundup; either way, price the 24/7 running cost in the power calculator first.

What to buy

What fits the advice above. Amazon links open a search, so you can compare current models and prices.

  • GMKtec EVO-X2

    The proven Strix Halo box for local AI: a Ryzen AI Max+ 395 with up to 128GB of unified memory, the unit most LLM benchmarks run on, and usually in stock.

    See it on Amazon
  • Minisforum MS-S1 Max

    The homelab pick: same 128GB chip with dual 10GbE, a PCIe slot for a NIC or HBA, and an internal power supply.

    See it on Amazon
  • Bosgame M5

    Usually the cheapest way to the full 128GB of unified memory, with a 2TB SSD, when you do not need 10GbE or a slot.

    See it on Amazon

FAQ

What is Strix Halo?

Strix Halo is the codename for AMD's Ryzen AI Max line, led by the Ryzen AI Max+ 395. It pairs up to 16 Zen 5 CPU cores with a large integrated Radeon 8060S GPU (40 RDNA 3.5 compute units) and a 50-TOPS XDNA 2 NPU, all sharing one pool of LPDDR5X memory. The point of interest for homelabs is that the pool scales to 128GB, so the GPU can address far more memory than any consumer graphics card, which is what lets these small boxes run very large AI models.

What is the best Strix Halo mini PC?

For local AI and value, the GMKtec EVO-X2: it ships the full Ryzen AI Max+ 395, takes 128GB, and is the unit most benchmarks were run on, usually in stock and usually cheaper than the workstation boxes. Get the Framework Desktop instead if you run Linux and want a repairable, reusable board, or the Minisforum MS-S1 Max if you want dual 10GbE and a PCIe slot for a homelab. The Bosgame M5 is the cheapest way to 128GB, and the HP Z2 Mini G1a is the pick when you need a vendor warranty.

How much does a Strix Halo mini PC cost?

More than it did at launch. A 64GB box starts around $1,800 and 128GB configurations run from roughly $2,000 to $3,500 or higher, depending on storage and the model. The jump is mostly the 2025 to 2026 memory price surge, since these use large amounts of LPDDR5X. Check current listings before buying, and if you do not need the full 128GB, a 64GB trim saves real money.

Can a Strix Halo mini PC run local LLMs?

Yes, and that is the reason to buy one. With 128GB of unified memory a single box can hold a 70-billion-parameter model at a 4-bit quant, or a large mixture-of-experts model like gpt-oss-120b, which will not fit on a 16GB or 24GB consumer GPU. You run them through Ollama, LM Studio, or llama.cpp. The limit is speed, not capacity: memory bandwidth around 256 GB/s means dense large models generate at a usable but modest pace, while MoE models run noticeably faster.

How big a model can 128GB of unified memory run?

Roughly a 70B model at 4-bit fits in about 40 to 48GB including context, so it runs comfortably, and you can go larger with mixture-of-experts models where only a slice of the parameters is active per token. On Windows you can usually allocate up to about 96GB of the 128GB as graphics memory; Linux can hand the GPU even more dynamically. Leave headroom for the operating system and any other services sharing the pool.

Does Strix Halo support CUDA?

No. CUDA is NVIDIA-only. Strix Halo uses AMD's Radeon iGPU, so local AI runs on ROCm, Vulkan, or llama.cpp instead. Ollama and LM Studio both support these paths and work well on the platform today, but the AMD software stack is younger than NVIDIA's, so expect to read a driver or build note now and then. If a tool you depend on is CUDA-only, this is the wrong box.

Can you upgrade the RAM on a Strix Halo mini PC?

No. The LPDDR5X is soldered on-package to hit the bandwidth the design needs, so there are no SODIMM slots. You choose 32GB, 64GB, or 128GB at purchase and that is permanent. Because the memory is the whole reason to buy one, size it for the models you intend to run and buy up if you are unsure, since you cannot add more later.

Is a Strix Halo mini PC good for gaming?

Yes. The Radeon 8060S is strong for an integrated GPU and plays most titles well at 1080p and many at 1440p, so these double as capable small gaming PCs. For a homelab, though, gaming is a side benefit; the reason to spend on Strix Halo over a cheaper mini PC is the large unified memory for local AI. If you only want a general home server, a used enterprise micro is far cheaper.

Strix Halo or a discrete GPU for local AI?

It is capacity versus speed. A Strix Halo box with 128GB can load models a single 16GB or 24GB card cannot hold at all, in a quiet unit that pulls under 150 watts. A discrete GPU has far more memory bandwidth, so it generates tokens several times faster on any model small enough to fit its VRAM. Buy Strix Halo to run big models slowly on little power; build a GPU rig to run smaller models fast. For the GPU route, start from [Ollama vs LM Studio](/compare/ollama-vs-lm-studio).

Related on HomelabCompass

← All self-hosted alternatives