The best PC for local AI on Linux in 2026: pick the driver stack, then the box

Every runtime that matters for local AI, from llama.cpp and Ollama to vLLM and ROCm itself, ships for Linux first and everywhere else later. Ubuntu 26.04 put NVIDIA’s CUDA toolkit and AMD’s ROCm platform into its own package archive in April. A Linux box can run a model with no desktop session eating graphics memory, and it can drive four cards where Windows Subsystem for Linux exposes one. On the software side, the argument is over: this is the best operating system for running models at home.

The hardware side is where it gets driver-shaped. A Windows buyer picks a memory size and any card with that number works. A Linux buyer picks a driver stack first, because the stack decides which cards are first-class, which kernel the distro must ship, and how much of your weekend the machine will ask for. NVIDIA is the safe path. AMD is excellent inside ROCm’s supported list and a manual override outside it. Strix Halo, the 128 GB mini PC chip, wants a kernel newer than most long-term-support installs carry. Intel is the cheapest 32 GB you can buy, wrapped in the most software. Choose the stack you are willing to maintain, and the box follows.

This is the Linux sibling of the Windows guide, and it uses the same memory logic as the RAM and VRAM post. Every price and version below was read from a vendor page, an official doc, or a price tracker on September 15 or 16, 2026, and is linked where it appears. Vendors and projects are compared on published facts only.

On Linux you choose a driver stack before you choose a box

A model runs on three layers you never think about on other systems. The kernel has to know the graphics chip exists; on Linux that is the in-tree amdgpu driver for AMD and Intel’s i915 or xe driver, or NVIDIA’s separately installed module. Above that sits the compute library the runtime calls into: CUDA for NVIDIA, ROCm for AMD, oneAPI for Intel, or Vulkan for any of them. Above that sits the runtime itself, llama.cpp, Ollama, or an app that bundles one.

On Windows the vendor hides those layers behind one installer. On Linux each layer is a decision with a version number, and the version numbers are where buying mistakes happen. A new Radeon card on a two-year-old kernel does not light up. A Strix Halo box on the wrong kernel shows ROCm a sixth of its memory. An Intel card without the vendor toolkit runs at half speed. None of those show up on a spec sheet.

So the honest order is stack, then card, then box. The hero above is the guide in one panel: three stacks, each with a kernel floor, a driver, a compute layer, and an upkeep cost, standing on one shared Vulkan floor.

NVIDIA is the safe path, and the open kernel module made it safer

For a decade the one tax on NVIDIA under Linux was the proprietary driver: a binary blob rebuilt against every kernel update, prone to breaking on a distro upgrade. That tax is mostly gone. NVIDIA’s open kernel modules became the default install with the R560 driver, cover every card from Turing onward with what NVIDIA calls equivalent or better performance, and are mandatory on Blackwell, which means every RTX 50 card. The user-space CUDA libraries are still closed, but the piece that used to break, the kernel module, is now open and packaged by the distros.

The install is one command. Ubuntu’s driver guide recommends sudo ubuntu-drivers install for a desktop and the --gpgpu flag for a compute box, and it recommends the tool specifically because it installs pre-built signed modules that work under Secure Boot. NVIDIA’s CUDA 13.4 installation guide supports Ubuntu 22.04, 24.04, and 26.04, Debian 12 and 13, Fedora 44, and RHEL 8 through 10. Ubuntu 26.04 is, per Canonical’s April 23 announcement, the first release to carry CUDA in its own repositories, and Arch ships CUDA 13.4.1 in its main repo. If you would rather not touch any of it, System76’s Pop!_OS ships a separate NVIDIA image for GTX 16 series cards and newer with the driver preinstalled. Ollama’s hardware page asks only for driver 550 or newer and compute capability 5.0, which reaches back to 2014 cards.

The catch with NVIDIA in 2026 is not software, it is the price of memory. The tracker at videocardprices.com had the 32 GB RTX 5090 at $6,899.99 on September 15, 245 percent over its $1,999 list, and the 16 GB 5070 Ti at $1,249.99. The used market has followed: a 24 GB RTX 3090, the card the build guide used to recommend as the bargain rung, was $1,400 used on September 16. NVIDIA is the low-upkeep stack. It is no longer the low-cost one.

AMD works inside ROCm’s list, and the list is the buying guide

AMD’s compute platform used to be the punchline of Linux AI. It is not anymore, but it remains a list, and buying a Radeon for local AI means reading the list before the price tag. AMD’s ROCm 7.2.1 Radeon support matrix names the whole RX 9000 line (9060, 9060 XT, 9070, 9070 XT, 9070 GRE), the top of RX 7000 (7900 XTX, 7900 XT, 7900 GRE, 7800 XT, 7700, 7700 XT), the PRO W7000 cards, and the Radeon AI PRO R9600 and R9700. It pairs them with Ubuntu 24.04.4 on the 6.17 hardware-enablement kernel, Ubuntu 22.04.5, and RHEL 10.1. The RX 7600 and the entire RX 6000 generation are not on it.

Two things soften the list. Ollama bundles its own ROCm libraries, and its Linux list goes further than AMD’s: it adds the RX 7600 and 7600 XT, the RX 6800 through 6950 XT, and the Ryzen AI Max+ 395. And for anything off both lists there is the HSA_OVERRIDE_GFX_VERSION variable, which tells ROCm to treat your chip as a similar supported one. llama.cpp’s build docs note it works on Linux and not on Windows, which is a small example of the whole thesis: the workaround exists here and nowhere else. It is still a workaround. A card on AMD’s own list is the one to buy.

The kernel is the other half of the AMD story, because the driver lives inside it. Phoronix’s RX 9070 launch review wanted Linux 6.12 at minimum, 6.13 preferred, with Mesa 25.0, and Ubuntu 22.04’s newest hardware-enablement kernel is 6.8. So an RX 9000 card on 22.04 is a paperweight, while the same card on 24.04 with the 6.17 HWE kernel, on 26.04’s 7.0, or on Fedora 44’s 6.19 is fine. Ubuntu 26.04’s release notes add ROCm 7.1.0 to the archive, and Arch carries rocm-hip-sdk 7.2.4, so on either you can skip AMD’s amdgpu-install package entirely. Either way, add yourself to the render and video groups or the runtime will see no GPU at all.

What you get for the reading is the only 24 GB and 32 GB cards that cost less than NVIDIA’s. The RX 9070 XT was $799.51 on September 15 against a $599 list, the 24 GB RX 7900 XTX was $1,348 new and $900 used on the 16th, and the 32 GB Radeon AI PRO R9700, a 9070 XT with double the memory, was $1,699.99 on Amazon the same day, a quarter of a 5090. Every one of those is on AMD’s list. That is the AMD deal on Linux: the memory is cheaper, and you pay for it in one afternoon of reading.

Strix Halo is the 96 GB rung, and it needs a newer kernel than most LTS installs ship

Past 32 GB, a Linux desktop without a five-figure card means the same chip the Windows guide leaned on: AMD’s Ryzen AI Max+ 395, where the processor and a 40-unit Radeon 8060S share up to 128 GB of soldered memory. The difference on Linux is what you have to do to reach it. There is no Adrenalin slider. AMD’s RDNA 3.5 system optimization page says to keep the BIOS reservation small, around 0.5 GB, and let the kernel’s shared pool do the work: the GTT pool defaults to about half of system RAM and can be raised with AMD’s amd-ttm helper. On a 128 GB box that lands you at the same 96 GB ceiling Windows gets, with the split under your control.

The kernel floor is the part people trip on. Until a fix landed in kernel 6.16.9, ROCm on Strix Halo saw roughly 15.5 GB of the memory the kernel had already allocated, and owners were patching boot parameters to get around it. AMD’s page now sets explicit minimums: Ubuntu 24.04 needs HWE build 6.17.0-19 or newer, its OEM kernel needs 6.14.0-1018 or newer, and other distributions need 6.18.4 or newer, with Fedora 43, Ubuntu 26.04, and Arch named as supporting it natively. Framework’s own Linux page for the Desktop is looser, recommending 6.15 or newer with Fedora 43, Ubuntu 25.10, and Bazzite as officially supported, but that floor is for the machine to work as a computer. AMD’s floor is for the GPU to see all of its memory. Take the higher number.

Cleared of that, the box performs as the Windows numbers suggest, and one Linux quirk works in your favor. An August test on a Strix Halo mini PC found llama.cpp on Vulkan generating faster than ROCm, 97.7 tokens per second against 73.7 on a 30B mixture-of-experts model, while ROCm was 21 percent quicker at reading the prompt. So even on the AMD-native path, the vendor-neutral backend is the one to try first for chat. For a dense 70B model, expect the pace StorageReview measured on HP’s Z2 Mini G1a: 4.24 tokens per second, and close to 40 on gpt-oss-120b, whose design only touches a few billion parameters per token.

Three boxes carry the chip and treat Linux as a first-class citizen. Framework’s Desktop is the community favorite, and its configurator had the 128 GB model at $3,449 and out of stock on September 16, with a 64 GB model at $1,959 and no operating system charge if you bring your own. A 192 GB version on the Ryzen AI Max+ Pro 495 was announced in July with no price or date. GMKtec’s EVO-X2 listed at $2,199.99 on sale from $2,599.99 with Ubuntu as a shipping option. And HP’s Z2 Mini G1a is Ubuntu-certified for 24.04, with the review unit above priced at $3,342.65 at B&H. Buy the 128 GB configuration in every case; the 64 GB box hands its GPU about 48 GB, which holds a 70B model with nothing to spare.

Intel Arc is the value wildcard: 32 GB at a 16 GB price, with a software tax

Intel’s Arc Pro B70 is the most interesting card in this guide and the one that best illustrates its thesis. It launched on March 25 at $949 with 32 GB of GDDR6 at 608 GB/s and a 230 W board, aimed squarely at local inference rather than games. The memory shortage has pushed it to $1,300 as of September 16, which is still the cheapest new 32 GB card on the market, roughly the price of a 16 GB RTX 5070 Ti and under a fifth of a 5090.

The tax is the stack. Intel’s fast path is SYCL, reached through its IPEX-LLM project, and the llama.cpp quickstart wants the oneAPI Base Toolkit 2025.0, a conda environment, and a source /opt/intel/oneapi/setvars.sh before every run. The B-series guide adds a driver PPA and insists Resizable BAR is on in firmware. Phoronix reviewed the B70 in April on Ubuntu 26.04 with kernel 7.0 and Mesa 26.0, which is a fair signal of the kind of distro that keeps this card happy. Miss any of it and you fall back to Vulkan, and here that costs real speed: one June test of Qwen 3.6 27B at 4-bit had SYCL at about 22 tokens per second and Vulkan at about 14. NVIDIA and AMD owners can ignore their vendor toolkit and lose a little. Intel owners cannot.

The reward for paying the tax is the cheapest path to 128 GB of discrete memory that exists. Puget Systems put four B70s in one Ubuntu 25.04 box under Intel’s vLLM container and ran Qwen 3.6 27B at 13.1 tokens per second across all four, with 8B models doubling from 35 to 70 tokens per second as cards were added. Four cards at today’s price is about $5,200 of GPU, less than one RTX 5090. It is a project, not a purchase.

Vulkan runs on all three, and costs 10 to 35 percent

Everything above has a fallback, and it is the same one. Vulkan is the cross-vendor graphics interface that llama.cpp can also use for compute with a single build flag, and Ollama’s hardware page says the backend is on by default on Linux. It needs no toolkit, no PPA, and no list. It runs on the RX 6700 that ROCm ignores, on an Intel card with no oneAPI, and on an NVIDIA card whose CUDA install went sideways. The one thing to know is that Ollama needs extra capabilities, or root, to read free memory through Vulkan for scheduling, which is a systemd unit edit rather than a blocker.

The cost is measurable and smaller than its reputation. On an RTX 5090, a spring comparison had CUDA generating 290 tokens per second to Vulkan’s 264, a 10 percent gap, while CUDA read prompts 36 percent faster. The same page’s Radeon numbers had Vulkan slightly ahead of ROCm at generation and behind on prompts, the Strix Halo test above found the same split, and only the Intel card opens a gap large enough to change a buying decision. The pattern is consistent: the vendor path wins at reading long prompts, Vulkan holds its own at writing the reply, and on AMD hardware Vulkan is a legitimate first choice rather than a fallback.

That table has one row worth being plain about. CSuite, the desktop app this blog belongs to, bundles Ollama, and its Linux build ships the Vulkan backend only: there is no CUDA runtime in the Linux package, so an NVIDIA card is accelerated through Vulkan there at the 10 percent cost above rather than through CUDA. The Linux build comes as an AppImage, which self-updates, and a Debian package, which does not. If you want the vendor path on Linux, run Ollama or llama.cpp directly, which the Linux apps roundup walks through.

Headless and multi-GPU are where Linux pulls ahead

Here is the part no Windows or Mac desktop offers. Ollama’s Linux install registers a systemd service that starts at boot, restarts itself, and runs under its own user. Edit it with sudo systemctl edit ollama, set OLLAMA_HOST=0.0.0.0, and the box serves every laptop on your network. Unplug the monitor and it keeps going. That turns the buying question inside out: the machine no longer has to be pleasant to sit at, it has to be quiet, cheap to run, and stuffed with memory.

Going headless also gives the model memory back. A desktop session is a GPU client: the compositor, the browser, and every Electron app hold allocations, and an April write-up on GPU memory pressure measured a single Electron app releasing about 300 MiB when closed, with GNOME and KDE at the heavy end and tiling window managers doing “almost nothing on their own.” A server install with no desktop hands the whole card to the model, and Ubuntu’s --gpgpu driver flag exists for exactly this machine.

Multi-GPU is the other exclusive. Microsoft’s WSL GPU page states that a machine with several GPUs can expose only one to WSL at a time. Native Linux exposes all of them, llama.cpp’s build docs let you pick devices with a --device flag and split a model across them, and vLLM does the same with tensor parallelism, which is how Puget ran a 27B model across four Intel cards. Two used RTX 3090s make a 48 GB pool for about $2,800 at the tracker price above, enough for a dense 70B model at 4-bit, which no single card under $10,000 holds.

Two cautions. The halves talk over the motherboard, so a board with two properly wired full-length slots matters more than the processor, and almost no prebuilt has one. And two 3090s draw 700 W under load before the rest of the box, which means a 1000 W supply and a case that moves air. If a single 32 GB card covers your ambition, buy that.

What each memory tier runs, and who sells it with Linux installed

The memory ladder is the same one every guide on this site uses, and the calculator will map your own model wishlist onto it. What changes on Linux is the vendor column: who will sell you the rung with the operating system already on it and supported.

System76. The Thelio Mira, redesigned in April, is the reference Linux tower. Its configurator starts at $1,699 and runs from an RTX 5050 through the RX 9070 XT, the 5070 Ti, the 5080, and the 5090, swapping in a 1000 W supply for the top two, with up to 192 GB of RAM and a choice of Pop!_OS 24.04, Ubuntu 24.04, or Ubuntu 26.04. Its Thelio Major workstation starts at $6,999 and reaches the 96 GB RTX PRO 6000 Blackwell with 256 GB of ECC memory. Every card on that menu is on a first-class stack.

Slimbook and TUXEDO. Slimbook’s Kymera Woodline configures up to an RX 9000 or RTX 50 card and 192 GB of RAM, and ships any of thirteen distributions from Ubuntu and Fedora to EndeavourOS. Germany’s TUXEDO Computers builds mini, midi, and maxi towers on AMD or Intel with its own Ubuntu-based TUXEDO OS preinstalled. All three guarantee the driver stack for the parts they sell, and on Linux that is worth paying for.

NVIDIA. DGX Spark is the one box in the guide that only runs Linux. NVIDIA’s spec page lists a Grace Blackwell GB10 chip, 128 GB of unified memory at 273 GB/s, 4 TB of storage, and inference on models up to 200 billion parameters, and its DGX OS is, per NVIDIA’s own guide, Ubuntu 24.04 with the drivers and CUDA preinstalled. The bandwidth is Strix Halo class, so a dense 70B model runs at Strix Halo speed. What you pay about $4,700 for, after the February price change NVIDIA staff confirmed on its forum, is the CUDA toolchain on an Arm desktop, which matters if you fine-tune and does not if you chat.

The shortlist at three budgets

These are stacks and card tiers rather than exact SKUs, because the SKU in stock next week will change and the stack will not.

Two things held across every page read for this post. First, the two AMD parts that need a kernel floor, RX 9000 and Strix Halo, both clear it on Ubuntu 26.04, Fedora 44, and a rolling distro, and neither clears it on Ubuntu 22.04. If you install anything older than 24.04 with the newest hardware-enablement kernel, buy NVIDIA. Second, the cheapest memory in every tier is on the stack with the most upkeep: Intel at 32 GB, used NVIDIA at 24 and 48, Strix Halo at 96. Decide how many evenings you will give the machine, pick the column that matches, and the box picks itself.

If the desk is a Mac, or you need this to travel, the Mac guide and the laptop guide run the same memory ladder without the driver question.