Framework Desktop: A 4.5-Litre Mainboard Built for Local AI

Framework made its name on laptops that can be opened, repaired and upgraded, so its first desktop was always going to be judged against that reputation rather than against a generic mini PC. The Framework Desktop is a 4.5-litre machine built around a custom mini-ITX mainboard, small enough to sit under a monitor and, by the company's own framing, aimed at two audiences at once: people who want a very small gaming PC and people who want to run large language models at home ^¹.

The AI case rests on one architectural decision. Framework is among the first to ship AMD's Ryzen AI Max processors, an all-in-one package that combines CPU, graphics and a neural processing unit with memory on the same package. That matters because the memory is shared rather than partitioned into a fixed pool of graphics VRAM. On a conventional desktop, the size of the model you can load is capped by how much VRAM the graphics card carries. On the Framework Desktop, the graphics cores can draw on a much larger unified pool, which is what allows unusually large models to run without buying a separate high-VRAM GPU.

The clearest illustration of that is the new top configuration. Framework has opened preorders for a 192GB version built on AMD's Ryzen AI Max+ PRO 495, with 16 Zen 5 cores, Radeon 8065S integrated graphics with 40 compute units, and 192GB of soldered LPDDR5X-8533 on a 256-bit interface. Framework quotes 273GB/s of memory bandwidth, up from 128GB of LPDDR5X-8000 in the earlier high-capacity model, and says both CPU and GPU clocks are slightly higher than the original Ryzen AI Max+ 395 configuration. The DIY Edition starts at $6,799, with initial systems scheduled to ship in November 2026 ^². Framework names quantized versions of DeepSeek-V4.1-Flash and MiMo V2.6 Flash RL as examples of models that can run locally, while cautioning that actual performance depends on the inference engine, quantization format, context size and workload.

That caveat is the honest core of the pitch. Fitting a model into memory is a necessary condition for running it locally, not a promise of speed. The Guru3D report makes the point directly: unified LPDDR5X is not HBM, and once a model fits, compute throughput can still be the limiting factor. Anyone who has read our breakdown of local AI models that fit in 16GB of system RAM will recognise the pattern — capacity decides what loads, and the silicon decides how pleasant the result is.

What the reviews actually say

The most detailed published assessment available is a review of the 64GB, Ryzen AI Max Plus 395 configuration, priced as reviewed at $1,987 with a 1TB WD_BLACK SN7100 boot drive, 5GbE, Wi-Fi 7 and Bluetooth 5.4 ^³. That review's verdict is notably mixed. It credits the machine with surprising performance and, more unusually for a 4.5-litre chassis, surprising sustained performance, and says the price starts to make sense once both are taken into account. But it also concludes that the Desktop gives up some of the user upgradability that desktop PCs generally, and Framework specifically, are known for — a direct consequence of integrating CPU, GPU, NPU and memory into a single package. The reviewer's view is that the memory reallocation advantage is realised mainly in AI workloads, and that for everyone else a mini-ITX gaming build, a full-size desktop or a cheaper mini PC may be the more effective buy.

That is a fair summary of the trade-off, and it should be read as one reviewer's assessment rather than a measured benchmark. No source here provides token-per-second figures for any model on either configuration, so the speed question — the one that decides whether a local model is usable or merely loadable — could not be confirmed. Nor do the sources establish how the 192GB machine performs under sustained inference, what it draws from the wall, or how loud it gets when the GPU is fully committed to a model.

The physical design is more straightforward. The case is 4.5 litres, smaller than a PlayStation 5 or an Xbox Series X, with a front panel made of 21 interchangeable plastic tiles that buyers configure when ordering. Framework's expansion cards appear at the front and back, two each, and can be specified as USB-C, USB-A, a headphone jack, an SD reader or a storage expansion card. Inside there is the mainboard, a fan, a heatsink, a power supply and two M.2 2280 NVMe slots. The newer mainboard keeps two M.2 positions and adds an open-ended PCIe 4.0 x4 slot, which physically accepts full-length x8 or x16 cards — useful for a capture card or extra networking, less so for a second GPU, since the slot's electrical width is fixed at four lanes.

Who it is for

Read the evidence together and the Framework Desktop looks less like a general-purpose desktop and more like a specific tool with a general-purpose case around it. If your goal is running models too large for a consumer graphics card's VRAM, the unified-memory design is the point, and the 192GB option extends that ceiling considerably — at a price that puts it in a different category from the $1,987 64GB unit. If your goal is gaming or a machine you will upgrade piece by piece over five years, the review's conclusion is that a conventional build or a cheaper mini PC makes more sense, and the soldered memory means the mainboard's useful life is fixed at purchase.

The question our earlier comparison of cloud AI subscriptions versus a local model raises is also the one the sources leave open: whether the workload justifies the hardware. A $6,799 desktop is a capital purchase against an ongoing subscription, and nothing in the supplied material quantifies the crossover. Nor is there published data on how a 192GB machine compares with a multi-GPU workstation for the same models.

What can be said with confidence is narrower. Framework has built a very small machine whose memory architecture is genuinely suited to local inference, and it has now announced a configuration with enough unified memory to hold models that would not otherwise fit on a desktop. What remains unestablished is the part buyers most want: how fast those models actually run, how the machine behaves under long inference loads, and whether the premium over a conventional build is recovered in practice. Until independent testing answers that, the Framework Desktop is a well-documented bet on memory capacity — not yet a demonstrated one on throughput.

Sources

1. ¹ Framework’s first desktop PC is optimized for gaming and local AI inference | TechCrunch - Framework, the company that is better known for its modular, repairable laptops, just released its first desk techcrunch.com
2. ² Framework Desktop Adds Ryzen AI Max+ PRO 495 and 192GB LPDDR5X for $6,799 www.guru3d.com
3. ³ Framework Desktop Review: Small and Mighty, but Shy of Upgrade Greatness tech.yahoo.com