Framework Desktop (Ryzen AI Max+ 395)

The Framework Desktop takes the interesting bit of the Strix Halo chip - up to 128GB of memory the graphics side can treat as its own - and drops it into a proper little x86 desktop you can actually tinker with. For anyone who wants big-model headroom without leaving Windows or Linux behind, this is the friendly option.

The take

This is the value pick of the unified-memory crowd, and it’s not close. For roughly half the price of a DGX Spark or a 128GB Mac Studio you get the same headline 128GB to load models into, on hardware that runs everything x86 does. It isn’t the fastest of the bunch - the memory tops out around 256 GB/s, so it shares the same bandwidth ceiling as the Spark - but for the money, the amount of model you can load is frankly brilliant. If you want the most capacity per pound and you don’t fancy learning a new operating system to get it, start here.

What it’ll actually run

The Radeon 8060S iGPU borrows from that shared pool, so with the memory split set generously in the BIOS you can hand it most of the 128GB. That means a 70B model at Q4 fits with room to spare, and you can sit comfortably in the 30B to 70B range that most people top out wanting at home. Smaller models in the 8B to 14B bracket run nicely and leave you plenty of memory for context.

Speed is the caveat, same as its rivals. At around 256 GB/s the bandwidth is the bottleneck, so a large model answers at a measured pace rather than instantly - expect single figures to low double figures of tokens per second on a 70B, quicker as you drop to smaller models. There’s a 50 TOPS NPU on board too, though for local LLMs it’s not much use yet - no inference tooling had made proper use of it at the time of writing, so the graphics side does the actual work. The software picture is the reassuring part: it’s plain x86, so Windows and Linux both run, and the ROCm stack keeps improving for the AMD graphics side. You will need to nudge the memory allocation in the BIOS to get the most out of it, but that’s a one-time bit of fiddling.

Who should buy it

Buy it if value and flexibility are what you’re after. It’s the machine for the tinkerer who wants a real desktop that runs anything, loads big models, and doesn’t cost flagship money to do it. The Framework side of things means it’s more serviceable than most mini machines, too, which suits this site’s crowd nicely.

Look elsewhere if you need raw speed or CUDA. If your models fit in 24GB, a used 3090 will be faster for less, and if you’re wedded to Nvidia’s tooling this isn’t your box. But for the most model you can load without spending silly money, on an OS you already know, the Framework Desktop is hard to beat.