Best Mini PC for Local AI and LLMs (2026)

Disclosure: Some links in this article are affiliate links. We may earn a commission on qualifying purchases at no extra cost to you.

The best mini PC for local AI is the GMKtec EVO-X2 in its 128 GB configuration, because 256-bit LPDDR5X on a 256 GB/s-class bus is the only memory subsystem on this list that keeps a 70B dense model, or a 100B-class mixture-of-experts model, resident in GPU-addressable memory with the bandwidth to run it. That ceiling is worth paying for when the same machine also serves a homelab; for a single user chatting with 7B to 14B models, the DDR5 picks further down do the job with a much smaller platform.

Two numbers decide what a box can actually do, and neither is an NPU rating. The memory ceiling decides which models load at all. The memory bus decides how fast tokens come out, because generating a token means reading the active weights once, all of them, every time. A 128 GB machine moving data at 89.6 GB/s is not the same purchase as a 128 GB machine moving it at 256 GB/s, even when the capacity line on the spec sheet reads identically.

Three hardware classes follow from that. Unified-memory Strix Halo systems reach 128 GB on a 256-bit bus. DDR5 SODIMM machines reach 96 GB or more on a 128-bit bus with both slots populated. Discrete-GPU mini PCs are quick for whatever fits in their VRAM, and the one built-in discrete GPU on this list has 8 GB.

As an Amazon Associate, this site earns from qualifying purchases.

Quick Picks

What Actually Decides LLM Speed and Model Fit

Decode is a memory read, and prompt processing is not

Every token passes the active weights through the memory controller once. Decode therefore scales with bus width and transfer rate before it scales with anything else, which is why a 40-CU integrated GPU on a 256-bit bus sits in a different class from a smaller part with a nicer FP16 number. The bandwidth figures in this article are bus arithmetic (bus width in bits multiplied by the transfer rate, divided by eight), and they are ceilings. Real decode efficiency lands below them, and the gap depends on the runtime: the same chip can look very different under Ollama and under llama.cpp.

Prompt ingestion is the other half of the experience and it is compute-bound, running on the same integrated GPU. A box built to stream weights will look ordinary when it has to ingest a 30,000-token document or answer several users at once. If prompt-heavy work is the point, a discrete GPU with enough VRAM fits the shape of the problem better than a bigger unified-memory box.

Capacity decides what loads

Leave room for the operating system before you count weights against the ceiling; a few gigabytes is a common working assumption on a Linux box that also runs a desktop or a hypervisor. On 128 GB that still leaves the 235B-class mixture-of-experts models reachable, but with little left over. Nothing above 128 GB fits a single machine on this list, and the memory on every Strix Halo box here is soldered, so the configuration you buy is the configuration you keep.

The NPU contributes nothing

llama.cpp, Ollama and LM Studio run on the GPU or the CPU. None of them dispatches to the XDNA 2 engine on a Ryzen AI Max+ 395, or to the NPU on an HX 370, or to Intel's AI Boost. The 50-TOPS NPU rating and the 126-TOPS platform sum GMKtec quotes describe hardware that local inference stacks do not use. Buying for NPU TOPS buys nothing, and a machine with a slower NPU and a wider memory bus beats one with the reverse.

VRAM allocation on Linux is a configuration step

On Linux the iGPU can address memory beyond the BIOS carve-out through GTT, and AMD's own documentation puts the default at roughly half of system RAM. The usual setup is a small carve-out in firmware plus the ttm.pages_limit kernel parameter, which is expressed in pages rather than gigabytes, and toolboxes that wrap this cap GPU-addressable memory a little under the physical ceiling. Whoever sets this up should leave a few gigabytes for the OS. A 96 GB carve-out with no GTT tuning is one reason a machine appears to have far less usable GPU memory than its RAM suggests.

ROCm is gated by kernel and version

AMD's ROCm documentation states that stable gfx1151 support requires a recent kernel (Ubuntu 24.04 HWE 6.17.0-19.19 or later, the OEM 6.14.0-1018 line, Ubuntu 26.04, or any distribution on kernel 6.18.4 or later) together with ROCm 7.2.x, 7.14.x or 10.x. ROCm 7.1.x is not supported on these APUs. Vulkan through RADV is the fallback and, at shorter context, often the faster decode path on this silicon. Plan for a weekend of driver work if you want the AMD compute stack, and plan for nothing extra if you accept Vulkan.

#1 GMKtec EVO-X2 (128 GB): The Capacity Pick

The argument for this machine is one line long: 128 GB of GPU-addressable memory on a 256-bit bus, which lets it hold a 70B dense model and still leave room for context, on a 256-bit bus. Everything else is a compromise you accept alongside it. There is one 2.5GbE port and no second NIC, so a box that also wants to be a firewall or a segmented homelab node needs a switch and VLANs. GMKtec rates the chassis at 120 W sustained, and owners of Strix Halo systems in this class routinely report that the higher power modes are audible, and some add external fans. Treat the quiet mode as the mode you will actually live in.

Storage is listed as a primary M.2 2280 PCIe 4.0 slot with a second position for expansion, and the variant list on GMKtec's own page disagrees with itself on capacity, so confirm the drive configuration on the listing rather than on the marketing page. Memory is soldered in 64 GB or 128 GB steps. There is no upgrade later, which makes the 128 GB decision permanent and worth getting right the first time.

Where it holds up:

• 128 GB at 256 GB/s, shared with the two other Strix Halo picks, is the configuration on this list that runs 70B dense and 100B-class mixture-of-experts models from memory
• 40 RDNA 3.5 compute units give the integrated GPU real headroom for offload beyond generation
• Two USB4 ports and Wi-Fi 7 cover the desktop side without a dock

Where it does not:

• Single 2.5GbE port rules out a router or a segmented network host without extra hardware
• Sustained 120 W in a small chassis means the fan curve is part of the experience
• Soldered memory: no path from 64 GB to 128 GB after purchase

Skip it if you need 10GbE, need silence more than capacity, or plan to serve several users at once on long prompts.

Check the GMKtec EVO-X2 on Amazon

#2 Beelink GTR9 Pro: Same Silicon, Real Networking

Beelink sells this in a single memory configuration, 128 GB, and pairs it with dual 10GbE and a vapor chamber under two turbine fans. For a machine that has to move model weights across a network or feed a second box, that is the difference between a desktop and a node. Beelink lists a 3-year warranty.

The caveat is specific and it matters for a 24/7 deployment. Owners report the 10GbE ports dropping link or hanging under sustained load, with the discussion pointing at the Intel E610 controller and at a combination of BIOS, controller firmware and driver that worked for some units and not others. That is a user report rather than a confirmed defect, and the firmware state may have moved since. Verify the current release notes before you commit a cluster to these ports.

Where it holds up:

• Dual 10GbE with 128 GB on a 256-bit bus, in a chassis with an internal power supply
• Vapor chamber and two turbine fans, validated at 140 W sustained per Beelink
• Internal PSU, no external brick on the floor

Where it does not:

• The 10GbE controller is the single most-reported problem on the model
• One memory configuration only, and it is soldered
• No PCIe slot for a discrete card

Skip it if the machine will sit on a network you cannot afford to troubleshoot.

Check the Beelink GTR9 Pro on Amazon

#3 Minisforum MS-S1 Max: The One With a PCIe Slot

The slot is the reason to pick this over the other two Strix Halo machines. It is a full-length connector wired at PCIe 4.0 x4, which makes it a better fit for an HBA or a multiport NIC for the homelab half of the job than for a full gaming card. Reviewers note poor airflow for cards in that slot, and the 320 W internal supply is a limit as well, so plan the build around those constraints instead of discovering them.

The four TDP modes from 60 W to 130 W are the most useful control on any box here: a machine that idles at a low power ceiling and ramps for a batch job is easier to live with than one locked at 120 W. That heat still has to leave a compact chassis through two fans, and fan noise is the recurring complaint in published coverage of this unit. Supply is the other open question; Minisforum has shown it intermittently unavailable, and a 192 GB MS-S1 MAX-P495 variant exists.

Where it holds up:

• A PCIe slot on a 128 GB unified-memory machine, which nothing else here offers
• Four power profiles from 60 W to 130 W, useful for an always-on node
• Dual 10GbE on a Realtek RTL8127 controller, per Minisforum

Where it does not:

• x4 electrical link and tight card clearance limit what the slot is good for
• Fan noise under sustained load is the standing criticism
• Availability has been intermittent, and a bigger-memory variant sits above it

Skip it if silence is a hard requirement, or if you expected the slot to take a full gaming card.

Find the Minisforum MS-S1 Max on Amazon

#4 GEEKOM A9 Max: The Sensible First Ollama Box

Most people asking about a mini PC for AI are running a 7B or 14B model, and this is the tier where that stops being a compromise. Twelve Zen 5 cores, a 16-CU RDNA 3.5 iGPU, two SODIMM slots you can actually open, and dual 2.5GbE for the rest of the homelab. The 89.6 GB/s ceiling means a 14B model at Q4 loads and generates at a conversational pace, a 32B model loads but crawls, and a 70B model does not load at all.

One check before you run anything: single-channel operation halves the bus. At least one unit has been reported to ship with a single module installed, so confirm what is in the box and add a matched second SODIMM if it arrived with one. That is the cheapest performance change available on this machine. A 2026 refresh with a Ryzen AI 9 HX 470 exists under a separate listing; the architecture is the same story, the extra TOPS are not.

Where it holds up:

• Real 7-14B capability in a mid-tier box, with room to add memory later
• Upgradeable SODIMM memory, unlike every Strix Halo pick above
• Dual 2.5GbE makes it a usable homelab node as well as a chat box

Where it does not:

• 89.6 GB/s caps the useful model size well below what the RAM slot count suggests
• Single-module configurations need fixing before the first run
• 32B models load but decode slowly at this bandwidth

Skip it if your target is a 30B model at chat speed, or anything at 70B.

Check the GEEKOM A9 Max on Amazon

#5 Minisforum UM890 Pro: Cheap Now, eGPU Later

This is the upgrade-path pick. On its own it is a 7B to 14B machine with the same 89.6 GB/s ceiling as the GEEKOM, and there is no reason to expect more from the Radeon 780M. The reason it earns a slot is OCuLink: a PCIe 4.0 x4 link out to a dock and a real graphics card, where speed is then set by the card's VRAM rather than by the mini PC. A 24 GB card puts 30B-class models at Q4 in reach, weights sitting in VRAM where the x4 link stops mattering for single-user generation.

Two things to check before relying on that path. The Radeon 780M is gfx1103, which is outside the gfx1150 to gfx1152 range covered by AMD's RDNA 3.5 optimization documentation, so the ROCm story on this iGPU depends on overrides or on Vulkan. And the eGPU route costs a dock plus a card, so the total is no longer a mid-tier purchase. If the plan is iGPU-only forever, buy the GEEKOM instead and get more cores.

Where it holds up:

• OCuLink gives a genuine path from 14B today to 30B-class with a card
• 96 GB ceiling and dual 2.5GbE for the price of a mid-tier box
• Two M.2 slots plus two USB4 ports

Where it does not:

• The iGPU itself is the weakest link in the lineup for model size
• ROCm support on gfx1103 is not part of AMD's documented RDNA 3.5 path
• Needs a dock and a GPU before the OCuLink story pays off

Skip it if you will never add an external card, or if you want 70B on one box.

Find the Minisforum UM890 Pro on Amazon

#6 Minisforum AtomMan G7 Ti: CUDA in a Small Chassis

The only machine here where the software question disappears. CUDA is the default target for every runtime, so Ollama, LM Studio and llama.cpp work without a driver argument, a kernel parameter or a Vulkan fallback. That is worth something on a machine that has to be reliable at 2 a.m.

Then the 8 GB of VRAM sets the ceiling, and it sets it hard. A 7B or 8B model at Q4 fits with room for context. A 14B at Q4 does not fit and spills into system RAM, where the DDR5 bus takes over and generation slows sharply. Anything 30B and above runs mostly from system memory. The G7 Ti is a fast small-model box and a competent gaming machine, and it is not a big-model box. A sibling with an RTX 5070 Laptop carries the same 8 GB limit. Minisforum's own stress figures put the combined CPU and GPU load inside what the chassis can hold; noise under full load is the standing complaint.

Where it holds up:

• CUDA everywhere, no ROCm kernel lottery
• 24 cores alongside the GPU, useful for CPU-side preprocessing and non-LLM work
• Doubles as a gaming machine, which no other pick manages well

Where it does not:

• 8 GB VRAM caps useful model size at roughly 8B at Q4
• Spending on a premium box to run 8B models is a lopsided value
• Audible under sustained combined load

Skip it if your target is 30B or larger; a Strix Halo box holds those and this one does not.

Find the Minisforum AtomMan G7 Ti on Amazon

#7 ASUS NUC 15 Pro+: Small, Intel Toolchain

A 0.7-litre office box that happens to run small models, with Thunderbolt 4. If the machine lives on a desk next to a person rather than in a rack, that form factor is the feature: nothing else on this list disappears behind a monitor this easily.

The Intel path is SYCL and OpenVINO rather than CUDA or ROCm, which means fewer tutorials and more of your own reading. Treat model-size claims as untested on this exact part. Bandwidth lands in the same 89.6 GB/s class as the HX 370 boxes, which is the real constraint rather than the core count. Two DIMMs are mandatory for the iGPU to have its full bus.

Where it holds up:

• 0.7 L chassis, Thunderbolt 4 and dual SODIMM slots
• 96 GB ceiling covers 70B weights in RAM, even if decode is bandwidth-limited
• SYCL, OpenVINO and Vulkan are all available on the Arc iGPU

Where it does not:

• No LLM figure attributable to this specific processor
• SYCL and OpenVINO have fewer tutorials than CUDA or Vulkan on AMD
• Same 89.6 GB/s ceiling as the mid-tier AMD boxes

Skip it if LLM throughput is the purchase criterion, or if you want a documented benchmark trail before buying.

Check the ASUS NUC 15 Pro+ on Amazon

#8 Minisforum MS-01: Only If You Already Want the Homelab

Buy this for the networking and the slot, not for the models. Two 10G SFP+ ports, two 2.5GbE ports, three M.2 positions and a low-profile PCIe slot make it one of the better small Proxmox hosts on this site, which is why it keeps showing up in Proxmox builds. The Iris Xe iGPU is useless for inference, and the CPU-only path at DDR5 dual-channel bandwidth is slower than the Strix Halo picks and in the same class as the 8945HS box.

No attributable LLM figure exists for this machine. A low-profile 16 GB workstation card has been reported to fit the slot, which would put roughly 14B to 24B at Q4 in reach with VRAM to spare, but that is a build report rather than a documented configuration and the slot is commonly described as x8 electrical rather than x16. Treat LLM work here as a secondary job that a card may or may not make comfortable. Supply is the first question anyway: Minisforum's store has shown the i9-13900H variants sold out.

Where it holds up:

• 10G SFP+ plus 2.5GbE plus a PCIe slot in one small chassis
• Three M.2 positions, one U.2-capable
• Established as a virtualization host

Where it does not:

• No usable iGPU for inference and no attributable LLM figure
• Low-profile single-slot only, with an electrical width that needs verifying
• Availability is uncertain and the model may be winding down

Skip it if LLM throughput is why you are shopping.

Check the Minisforum MS-01 barebone on Amazon

Comparison Table

Apple's Mac mini sits outside this table on purpose: the M5 Pro carries 307 GB/s per Apple's spec page and a 64 GB ceiling, which is a different shape of machine rather than another row.

The Mac mini question

The other option worth pricing against these is an Apple Mac mini with an M5 Pro. Apple lists 307 GB/s and a 64 GB ceiling, Metal and MLX run without driver work, and decode on small models sits comfortably with the best of this list. What it is less suited to is a 70B model: 40 to 43 GB of Q4 weights in 64 GB of unified memory leaves limited room for context and everything else, no measured 70B result is available for it, and there is no path past 64 GB on the mini. For one person chatting with 7B to 32B models, it is a serious candidate. For 128 GB capacity, dual 10GbE or upgradeable memory, it is not, and Apple's spec page is the place to confirm the current configuration rather than a reseller listing.

Check the Apple Mac mini M5 Pro on Amazon

Frequently Asked Questions

Conclusion

Memory bandwidth and capacity decide this market, and they split it cleanly in three. The 128 GB Strix Halo boxes are the only picks here with a memory bus wide enough to make 70B dense and 100B-class mixture-of-experts models workable, and among them the GMKtec EVO-X2 is the simplest expression of the idea, the Beelink GTR9 Pro adds real networking with a caveat attached, and the Minisforum MS-S1 Max is the one to take if you want a PCIe slot in the same chassis.

Below that line, the GEEKOM A9 Max is the machine most people should actually buy, because a 14B model at Q4 is what the majority of home use looks like. The UM890 Pro is the same tier with an exit to a real GPU through OCuLink. Above it and beside it, the G7 Ti is the CUDA answer for small models and the MS-01 is a homelab host that tolerates a little model work on the side.

Whichever way you go, buy the memory configuration you intend to keep: on every 128 GB and 64 GB Strix Halo machine here it is soldered to the board. If budget bands are the open question rather than silicon, the price guide covers what each tier currently buys, and the Beelink and Minisforum brand guides map the rest of each lineup. For setting the machine up once it arrives, the Ubuntu guide covers drivers and the bits that come before the first model load.

Check the GMKtec EVO-X2 on Amazon