The Framework Desktop crams workstation-class silicon into a 4.5-liter chassis that barely occupies more desk space than a large book. Mine is the top-end configuration: an AMD Ryzen AI Max+ 395 (Strix Halo) with 16 Zen 5 cores, 40 RDNA 3.5 compute units, and 128GB of unified LPDDR5X memory soldered to the package. It runs Linux exclusively and handles everything I throw at it as a daily driver. The expansion card system carried over from Framework’s laptops is a better idea on a laptop than a desktop, and the I/O is thinner than I’d like. But the build experience was genuinely enjoyable. The 128GB of unified memory means I can run 70-billion-parameter models that won’t fit on even a high-end discrete GPU—slowly, but they run. The PCIe slot is a baffling design miss, and Linux software support for the Strix Halo platform is still catching up to the hardware. But even with those caveats, this is the most compelling small form factor workstation I’ve used.

What It Is

The Framework Desktop is a compact, modular desktop PC built around AMD’s Strix Halo platform. The Ryzen AI Max+ 395 is the flagship SKU: 16 cores, 32 threads, boost to 5.1GHz, with 40 RDNA 3.5 integrated graphics compute units sharing the system’s memory pool. There’s no discrete GPU—the integrated graphics and the CPU draw from the same 128GB of LPDDR5X, which is soldered directly to the processor package for bandwidth.

Storage is two M.2 2280 NVMe slots running PCIe 4.0. There’s a single PCIe 4.0 x4 expansion slot inside. The front panel accepts Framework’s expansion card modules—the same ones used in their laptops—for configurable I/O. The whole thing measures roughly 97 x 206 x 226mm.

At $1,999 for the 128GB configuration before storage, cooling, and accessories, it’s not cheap. But it’s not priced unreasonably for what the Strix Halo platform delivers, either.

The Build

Framework markets the Desktop as a kit you build yourself, which is mildly deceptive—in a good way. The “build” consists of screwing in a fan, seating an NVMe drive, snapping in expansion cards, and attaching side panels and cosmetic tiles. Most of the choices are aesthetic. I’ve been building my own computers for thirty years, and this barely registered as a build. It was just Tuesday.

Everything fit perfectly. The tolerances are tight, the documentation is clear, and nothing about the process felt like a compromise for the sake of modularity. It’s nice, in the way that well-engineered things are nice—you appreciate the precision even when it doesn’t demand anything of you.

If the kit aspect scares you, don’t let it. If you can assemble IKEA furniture, you can build this. The only real decision is which color tiles you want on the front panel.

Living in 4.5 Liters

The form factor is the first thing that hits you. This machine has the same CPU that would comfortably anchor a full-tower workstation, and it sits on my desk like a small appliance. The power density is remarkable. There’s no wasted space inside, and the external dimensions mean it fits places where a traditional desktop simply wouldn’t.

For a daily driver workstation, the size has practical benefits beyond aesthetics. It’s easy to move, it doesn’t dominate a desk, and it runs quietly enough to forget it’s there during normal workloads.

Expansion Cards

Framework’s expansion card system is a genuinely revolutionary idea—on a laptop. My Framework Laptop 16 has six expansion card slots, which is enough to meaningfully customize your I/O for different workflows. On the Desktop, you get two. Two choose-your-own front ports that are essentially USB-C with a plastic support bracket instead of a dongle.

The rest of the Desktop’s I/O is fixed: two USB-C ports on the back, a pair of USB-A for your mouse and keyboard, and the Ethernet port. That’s it. The first accessory I bought was a seven-port USB hub, which tells you everything about how well the built-in port selection covers a real desktop workflow.

The expansion card concept is great, and the execution is great. There just aren’t enough of them to matter on a desktop. What was revolutionary on a laptop lands as merely adequate here—two swappable ports don’t transform the I/O experience the way six do.

The PCIe Slot

This is the one design decision that genuinely baffles me. The Framework Desktop includes a single PCIe 4.0 x4 slot, but it’s physically closed-ended—you can’t seat a longer card—and the chassis has no rear opening to expose external ports from whatever card you install. A 10GbE NIC, an Oculink adapter, a USB expansion card—any of these would physically fit in the slot, but there’s no way to route ports outside the case without taking a Dremel to it.

What were they thinking? If the form factor couldn’t accommodate a proper slot with a rear knockout panel, why not reallocate those PCIe lanes to something useful—another NVMe slot, a built-in 10GbE port, anything? Instead, the lanes sit behind a slot that’s essentially decorative for most use cases. The community has been vocal about this, and Framework has acknowledged the feedback, but it remains the most puzzling design choice on an otherwise thoughtfully engineered machine.

In a product that nails modularity everywhere else, this feels like someone added the slot to a spec sheet without thinking through whether anyone could actually use it.

5-Gigabit Ethernet

While we’re on the subject of puzzling I/O decisions: the Framework Desktop includes a single 5Gbit Ethernet port driven by a Realtek controller. It’s a strange middle ground. Most consumer hardware ships with 2.5GbE. Power users and homelab setups are moving to 10GbE. Five gigabit splits the difference in a way that satisfies neither camp—it’s more than most people’s switches can use, and less than the people who actually care about Ethernet speed want.

For a machine with a useless PCIe slot that could have hosted a 10GbE NIC, the choice is especially frustrating. Why not 10GbE on the board? Why not drop the PCIe slot entirely and use those lanes for a second Ethernet port, or another NVMe slot, or anything with a clear use case? The theme here is the same as the PCIe slot: the Framework Desktop makes a few I/O decisions that feel like they were made in isolation from how the machine would actually be used.

Linux on Strix Halo

I’m running this machine exclusively on Linux, and the experience has been… a journey. I tried Debian first—normally my go-to distribution—then NixOS as an experiment, and finally settled on Ubuntu, which is what’s running now. All three functioned for normal desktop use, but only Ubuntu had decent hardware support at release time, and even that was incomplete when I got mine early.

The situation has improved as kernel and Mesa updates have landed, but software support for the Strix Halo platform still isn’t quite there. It works well enough for daily use, but getting the right libraries installed—particularly for GPU-accelerated workloads—was tricky and still requires some workarounds. Framework has consistently prioritized Linux support across their product line, and I trust this will continue to improve. But if you’re buying one today expecting everything to work out of the box on your distribution of choice, temper that expectation. Ubuntu with a current kernel is your safest bet.

Local LLMs

I’ll be honest: the local AI experience is mixed, and the nuance matters.

The Strix Halo’s main selling point for AI workloads is the huge pool of unified memory, and that delivers exactly as advertised. I can load 70-billion-parameter models—models that are too large even for an RTX 5090’s VRAM. They load, they run, they produce output. In a 4.5-liter box on my desk. That’s genuinely impressive.

The large models run very slowly.

Getting there took some effort. After spending a few weeks sorting out hardware support and library dependencies for Ollama, the performance picture is more nuanced than “fast” or “slow.” On smaller models, the Framework Desktop is reasonably quick—it won’t compete with a dedicated graphics card, but it won’t embarrass itself either. On the massive models that are the whole reason you’d buy this configuration, inference is sluggish. The memory bandwidth of LPDDR5X, shared between CPU and GPU, simply can’t match dedicated GDDR or HBM. But small models aren’t why you buy this system.

What keeps the experience tolerable is the thermals. The hefty heatsink and fan keep the APU chugging along in reasonable quiet even under sustained AI workloads—no thermal throttling, no jet engine whine. It does the work without complaint; it just does it at its own pace.

There’s another catch that nobody mentions in the marketing: unified memory means shared memory, and the split is a BIOS setting, not a dynamic allocation. If you want to run large models, you set the GPU to claim 96GB of the 128GB pool—and that’s where it stays until you reboot and change the BIOS. You could switch it back, or split it 64/64 as a compromise, but if you’re not dedicating the full 96GB to the GPU, why did you buy the 128GB configuration? In practice, you pick a VRAM allocation and live with whatever’s left. For me, that means 32GB of system memory. Thirty-two gigabytes is enough—barely. I hit the OOM killer a few times before I manually added more swap space, and I had to be more deliberate about what I left running alongside normal desktop use. Most people would be fine with 32GB of headroom. Then again, most people aren’t buying the 128GB configuration to run 70-billion-parameter models. The Venn diagram of “people who buy this system” and “people who will hit this limit” is close to a circle.

The honest framing is this: you don’t buy a Framework Desktop because it’s fast at AI inference. You buy it because most dedicated graphics cards can’t run these models at all. A 70B model doesn’t fit in 24GB of VRAM, no matter how fast that VRAM is. The Framework Desktop with 128GB of unified memory runs it slowly, and you’ll want to keep an eye on what else is competing for that memory while it does. Everything else at this price point doesn’t run it. That’s the value proposition, and it’s a real one—just don’t mistake it for a speed advantage.

Verdict

The Framework Desktop is an opinionated machine, and its opinions are mostly right. The Strix Halo platform delivers workstation performance in a form factor that barely registers on a desk. The build experience respects your time and intelligence. And 128GB of unified memory lets you run models that won’t fit on anything else at this price.

It’s not without rough edges. The PCIe slot is bafflingly unusable. The expansion cards are a better idea on a laptop than on a desktop with only two slots. The 5GbE Ethernet is a puzzling choice. Linux support is solid but still catching up to the hardware—expect some library wrangling, especially for GPU-accelerated workloads. And local LLM inference is slow enough that you’ll want to set realistic expectations about throughput before buying this as an AI machine.

But none of that changes the core proposition. This is a compact, powerful, repairable workstation that runs Linux well, fits on a desk without dominating it, and gives you access to model sizes that are simply out of reach on most consumer hardware. If that’s what you need, the Framework Desktop with the Ryzen AI Max+ 395 and 128GB of memory delivers.