Best Mini PC for Local LLMs (2026): Strix Halo Boxes, DGX Spark-Class Systems, eGPU Minis and Jetson Compared

The best mini PC for a local LLM in November 2026 is a Ryzen AI Max+ 395 box with 64 or 128 GB of unified memory, from $2,199.99, because it holds large models in a quiet 4-litre case. Choose an NVIDIA GB10 system if you need CUDA, a Mac mini from $899 for small models, an OCuLink mini with an external GPU for speed, and a Jetson for edge projects.

A mini PC is an appealing home for a local model: it sits on a shelf, draws little power and can stay on. But the category now covers machines that differ more than tenfold in price and far more than that in what they can run. The useful question is not which box is fastest but which class matches the models you want and the software you already use.

Prices in this guide are US list prices read from each vendor’s own store, and specifications come from the manufacturers’ product pages, on 14 and 15 November 2026. Prices in this category have moved several times this year, so treat every figure as a snapshot. Speeds are bandwidth-based estimates from our tokens-per-second calculator unless a source is named.

Quick picks: best mini PC for local AI by need

What makes a mini PC good for LLMs

Three numbers decide what a small machine can do, and none of them is the processor’s core count.

1. Memory the GPU can reach. A model must fit in memory that the accelerator can address. In a unified-memory machine that is most of the system RAM. In a mini with a separate graphics card it is the card’s own memory. In a plain mini with integrated graphics it is system RAM, at system-RAM speed.
2. Memory bandwidth. Generating a token means reading the model’s active weights once, so bandwidth caps speed. Ordinary dual-channel DDR5-5600 delivers 89.6 GB/s by arithmetic. The larger unified-memory chips in this guide deliver 256 to 307 GB/s. A desktop graphics card delivers several times more.
3. Software stack. CUDA on NVIDIA, ROCm or Vulkan on AMD, Metal and MLX on Apple. This decides which tools work on day one and whether your set-up transfers to a server later.

Everything else, including NPU ratings, matters much less. Our laptop guide explains the bandwidth arithmetic in more detail.

AMD unified-memory mini PCs: models and prices

AMD’s Ryzen AI Max+ 395, known as Strix Halo, combines 16 Zen 5 cores, a 40-unit Radeon GPU and up to 128 GB of LPDDR5x-8000 on a 256-bit bus. AMD says a 128 GB system supports models of up to 200 billion parameters. Many vendors now build the same chip into small desktops, so the choice comes down to price, ports, cooling and support.

A few buying notes:

• 64 GB or 128 GB. The 64 GB EVO-X2 costs $1,300 less than the 128 GB one. It runs everything up to Qwen3.6-35B-A3B comfortably. Pay for 128 GB only if you want the 120B class.
• A 192 GB tier now exists. Framework lists a Desktop with the Ryzen AI Max+ PRO 495 and 192 GB of LPDDR5x-8533, and AMD’s Halo page shows the same chip, which AMD rates for models of up to 300 billion parameters. Framework’s own table pairs it with DeepSeek V4.1 Flash at a 2-bit quantization using 163 GB.
• Check GPU-addressable memory. HP states 96 GB of 128 GB; other vendors expose the setting in firmware. Ask before you size a model against the full 128 GB.
• Memory is soldered. None of these can be upgraded later.

For set-up, drivers and measured behaviour on this chip, see our separate Strix Halo guide.

DGX Spark and GB10 partner systems

NVIDIA’s DGX Spark is a 150 mm square box, 50.5 mm tall and 1.2 kg, built on the GB10 chip: 128 GB of unified memory at 273 GB/s, up to 4 TB of self-encrypting storage and a 240 W supply. NVIDIA says it runs inference on models of up to 200 billion parameters and fine-tunes models of up to 70 billion.

Buying one is less simple than it was. NVIDIA’s product page lists no price for the 128 GB Founders Edition, and its Marketplace listing would not load for us on 15 November 2026. In October NVIDIA announced a 64 GB configuration “starting at $4,999”, sold only through partners from 23 October, and its scaling table rates that version for models of up to 100 billion parameters. Partner machines on the same chip include the Acer Veriton GN100, ASUS Ascent GX10, Dell Pro Max, Gigabyte AI TOP ATOM and HP ZGX Nano, with Lenovo and MSI also listed. Prices differ by retailer, so compare the listing on the day.

The reason to pay more than a Ryzen box costs is software. GB10 runs NVIDIA’s Linux with CUDA, so serving engines such as vLLM run as they do on data-centre GPUs, and NVIDIA’s scaling table goes up to four linked units. Our desk-side hardware comparison weighs it against the alternatives in detail, and NVIDIA’s Windows sibling is covered in the RTX Spark guide.

The Apple option, in brief

Apple’s US store lists the Mac mini with M6 and 16 GB from $899, and with M5 Pro and 24 GB from $1,699, configurable to 64 GB. Apple gives the M6 up to 170 GB/s of memory bandwidth and the M5 Pro 307 GB/s; our MacBook Pro M6 guide shows what the same chip holds at each memory size. It is the quietest and cheapest way into the category, with the limits that it runs macOS only and stops at 64 GB. Model picks for each configuration are in our guide to local models on Apple hardware.

Mini PCs with OCuLink, an eGPU or a low-profile GPU

Unified memory buys capacity. A graphics card buys speed. If the models you want fit in 16 to 32 GB, a cheap mini with a real GPU attached will generate tokens several times faster than any box above.

OCuLink and an external dock. OCuLink carries PCIe out of the case. Minisforum’s X1 Lite, with a Ryzen 7 255, two DDR5-5600 slots for up to 128 GB, OCuLink and USB4, is listed at $311 as a barebone or $743 with 32 GB and 1 TB, though every configuration showed as sold out on 15 November. The DEG1 dock is $119, provides a PCIe 4.0 four-lane link and takes a standard ATX or SFX power supply. Add a desktop card and you have a compact machine whose model memory is the card’s. The link is narrower than a desktop slot, which mainly lengthens model loading, because generation happens inside the card’s own memory.

A low-profile card inside. A few minis have a PCIe slot. The cards worth knowing are the 70 W workstation models:

Intel specifies a 280 W minimum system supply for the Arc Pro B50. At 224 GB/s it should run gpt-oss-20b at about 37 tokens a second by our estimate. To choose a full-size card for a dock, use the GPU buyer’s guide and the model picks by graphics memory.

The trade-off is tidiness. A dock, a power supply and an exposed card are no longer a mini PC in any meaningful sense, and the fans on a desktop GPU are louder than any box in this guide.

Intel Core Ultra mini PCs

The ASUS NUC 16 Pro is the reference design for this class. ASUS lists it with Intel Core Ultra processors up to the X9 388H, Intel Arc integrated graphics, two Thunderbolt 4 ports and a 120 W or 150 W adapter, in a case of 144 by 117 by 42 mm. Memory is either two slots taking up to 128 GB of DDR5-7200 modules, or soldered LPDDR5x-8533 in 32 and 64 GB versions.

DDR5-7200 on two channels gives 115.2 GB/s by arithmetic, and LPDDR5x-8533 gives 136.5 GB/s. That is roughly half of a Strix Halo box, with a far smaller integrated GPU. In practice an Intel mini is an office PC that also runs models: fine for a 9B assistant or gpt-oss-20b at around 19 to 23 tokens a second by our estimate, slow for anything dense above 20B. The appeal is price, Windows manageability and slotted memory you can upgrade.

Jetson boards

NVIDIA’s Jetson line is built for devices rather than desks, but it is the smallest CUDA hardware you can buy. NVIDIA’s Jetson FAQ listed these developer kit prices on 15 November 2026, along with $3,499 for the older Jetson AGX Orin kit:

The Orin Nano Super suits 2B to 4B models. Google’s LiteRT-LM card for Gemma 4 E2B reports 24.2 tokens a second on its GPU and 12.2 on its CPU. The Thor kit has the same memory and bandwidth as a GB10 desktop; NVIDIA announced it at $3,499 in August 2025 and its FAQ now lists $5,499. Buy a Jetson when the model has to live inside a product or at a site with a tight power budget. Our article on CPU and NPU inference at the edge covers that deployment pattern.

What each mini PC can run

File sizes are the GGUF downloads published by the llama.cpp project and LM Studio on Hugging Face. Speeds are estimates for one user from the calculator’s method, which assumes a runtime reaches 60% of peak bandwidth; they are not benchmarks.

Two patterns matter more than any single number. Mixture-of-experts models are the right choice for every unified-memory box, because they read only a few billion parameters per token: gpt-oss-120b is about eight times faster than a dense 70B model on the same machine. And capacity without bandwidth disappoints: a 128 GB DDR5 mini can load the same file as a Strix Halo box and run it at a third of the speed. To test a specific model and window before buying, use the VRAM and KV cache calculator.

Mini PC for Ollama: what works

Ollama is the usual first install on any of these machines. From its documentation, as it applies here:

• Acceleration. NVIDIA GPUs through CUDA, AMD through ROCm version 7 with the Ryzen AI Max+ 395 on the supported list, Apple silicon through Metal, and other GPUs on Windows and Linux through Vulkan. With none of those, it runs on the CPU.
• Context. The default window follows GPU memory: 4K below 24 GiB, 32K from 24 to 48 GiB, 256K from 48 GiB. On a small mini, set it yourself.
• Network. The server binds to 127.0.0.1 on port 11434. To use the mini from other devices you must change OLLAMA_HOST, and the API has no login of its own, as our Ollama security review explains.

ollama run gpt-oss:20b
OLLAMA_CONTEXT_LENGTH=32768 ollama serve
ollama ps

A mini PC makes a good always-on endpoint for one person or a household. For a team sharing one box, read what changes with several Ollama users first: requests queue, and every concurrent session needs its own context memory.

Noise and power

Supply ratings are ceilings, not typical draw. AMD’s own cost example for a Ryzen AI Max+ system assumes 150 W sustained as a worst case, which it prices at $16.20 a month at $0.15 per kWh running all day. Only Framework publishes a noise figure, so for a machine that will sit beside you, look for an independent review of the exact model. Smaller cases with smaller fans are generally louder under sustained load.

Who should buy what

• A developer who wants large models locally: a 128 GB Ryzen AI Max+ 395 box, or 64 GB if 35B-class models are enough.
• A team that will deploy on NVIDIA servers: a GB10 system, so the serving stack you test is the one you ship.
• Someone who wants a silent always-on assistant: a Mac mini with 24 or 32 GB.
• Someone who values speed over capacity: an OCuLink mini and a 16 to 24 GB graphics card.
• An IT team standardising office hardware: an Intel Core Ultra mini with 32 GB or more, accepting small-model speeds.
• A product or site with a power limit: a Jetson.

How VDF AI fits

VDF AI does not sell, resell or ship on any of the mini PCs in this guide, and it has no integration with a particular box. It is a platform for organisations that run governed AI on their own servers, private cloud or an air-gapped network. The honest link is the step after the purchase: a mini PC serves one person well, and a department needs to know who hosts the model, which models are approved, who can reach them and what is recorded.

VDF AI Router registers Ollama and custom on-premises deployments alongside any cloud models your policy allows, and its air-gap mode restricts routing to local models. The self-hosting requirements describe what the platform itself needs. To discuss a pilot, contact us.