Framework Desktop (Ryzen AI Max+ 395) Review

The Framework Desktop is an unusual machine, and deliberately so. Rather than chase the enthusiast desktop market with slots and sockets, Framework has built a 4.5-liter box around AMD's Ryzen AI Max+ 395 — 16 Zen 5 cores, a 40-CU Radeon 8060S and an XDNA 2 NPU — and pointed it squarely at people who want to run large language models on their own desk. The headline trick is the memory: up to 128GB of unified LPDDR5X that the integrated GPU can draw on as a giant VRAM pool, which is something no conventional consumer GPU at this price can offer.

That makes the audience fairly specific. If you want to load a 70B-class model locally, experiment with quantised models far larger than a 24GB graphics card will hold, or simply keep sensitive inference off the cloud, this is one of the few $2,000 options that can even attempt the job. It's also a credible general-purpose small-form-factor PC — independent benchmark runs covering productivity and gaming workloads suggest it holds up well for its size — and it idles at around 12.5W, which matters for a machine you might leave serving models all day.

Less suited: anyone who needs CUDA, anyone planning serious training rather than inference, and anyone who expects a Framework product to be as upgradable as the company's laptops. Two of those three are hard limits.

The value argument is the strongest thing here. At $1,999 for a 128GB-class configuration, this reportedly lands at roughly half the cost of an NVIDIA DGX Spark or a similarly specified Mac Studio, and that gap is wide enough to forgive a lot. The DIY Edition's repairability focus, mini-ITX mainboard and customisation options (I/O choices, colored tiles) also give it a character that generic mini PCs lack, backed by the unusually thorough assembly documentation Framework is known for.

The compromises are real, though. Soldered memory is the one that stings most — it is flatly at odds with Framework's own upgradability pitch, and it means your purchase decision is permanent on the spec that matters most for AI work. The 256 GB/s of bandwidth is the other ceiling: it's what lets a 70B model fit, but it's also why you should expect single-digit to low double-digit tokens per second on one. That's fine for batch work, tinkering and overnight jobs; it is not a comfortable interactive chat experience. And without CUDA, you're on the ROCm/Vulkan path, which remains less polished and will cost you time on tooling that "just works" elsewhere.

Assembly is reportedly fiddlier than it should be — a beta driver bundle, an awkward top panel and an SSD install that reviewers found nerve-wracking are all avoidable rough edges on a product built around ease of servicing. The competitive picture has also tightened: with the 128GB configuration reportedly up around $460 in 2026 on RAM shortages, rival Strix Halo boxes such as the Beelink GTR9 Pro now sit at similar money, which erodes the price advantage that justified the other compromises.

I'd rate it well for what it uniquely does and dock it for the things it can't be talked out of. It's an excellent, genuinely interesting local-inference appliance and a good small desktop; it is not a flexible workstation, not a training box, and not upgradable. Buy it knowing exactly which of those you need. One caveat on confidence: the performance and pricing figures here come from third-party reporting rather than our own lab testing, so treat the token-rate expectations as directional.