Framework Desktop 192GB pre-orders open at $6,799

Framework opened pre-orders on September 30, 2026 for a Framework Desktop with 192GB of memory. It is built on AMD's new Ryzen AI Max+ PRO 495 chip, codenamed Gorgon Halo. The extra memory lets one small box hold open-weight AI models of more than 300 billion parameters. The Register puts the price at $6,799.

Framework says shipments start in November 2026. It will build only one batch at first, using the memory it has in stock.

What changes from the 128GB model

The memory is the headline. Framework's announcement puts it plainly: "The important change for local AI is the extra unified memory." Unified memory means the processor and the graphics chip share one pool of RAM. A separate graphics card, by contrast, has its own memory.

Framework now sells four memory sizes across three AMD chips.

Memory	Chip
32GB	Ryzen AI Max 385
64GB	Ryzen AI Max+ 395
128GB	Ryzen AI Max+ 395
192GB	Ryzen AI Max+ PRO 495

The new chip has 16 Zen 5 processor cores and a graphics chip with 40 compute units, Framework's product page says. Tom's Hardware reports that the Ryzen AI Max 400 family uses AMD's RDNA 3.5 graphics design and clocks up to 5.2 GHz. The memory is LPDDR5X-8533 on a 256-bit bus. Framework gives the bandwidth as 273GB/s, which is how fast data moves between memory and the chip.

Buyers can pick a DIY Edition to assemble themselves or a pre-built system that ships with Fedora Linux.

How big a model fits

The Register worked out what 192GB can hold. At 4-bit precision, it can run models of about 345 billion parameters on Linux and about 320 billion on Windows. Four-bit precision, also called quantization, stores each of the model's numbers in 4 bits instead of 16. That cuts memory use to about a quarter, at some cost in accuracy.

The Register names two open-weight models that fit: DeepSeek V4 Flash, with 284 billion parameters, and GLM-5.3-Flash, with 320 billion. Neither would load on the 128GB version at that precision.

Why the price is so high

Framework's blog says the 192GB model is a step up in cost over the 128GB one. The Register reports the $6,799 price and notes that "all that memory doesn't come cheap."

The whole category has grown more expensive, The Register reports. A year ago, machines built on AMD's previous Strix Halo chip cost $2,000 to $3,000. By summer 2026 they cost about $4,000. Systems built on Nvidia's GB10 chip once sold for $3,000 to $4,000 and now run $6,000 to $8,000.

What this means for developers

Buy this for model size, not for speed. The 192GB box exists to load models too big for the 128GB version. How fast it answers depends largely on memory bandwidth, and 273GB/s is the number to plan around. Before you order, estimate the tokens per second you need and compare that with published tests of similar machines.

Leave room beyond the model itself. A 320-billion-parameter model that "fits" at 4-bit still needs memory for its context, the running record of the conversation. Long prompts and long documents eat into that space. Pick a model with a comfortable margin, not one that fills all 192GB.

Plan on Linux if you want the largest models. The Register's figures give Linux about 25 billion more parameters of headroom than Windows. The Fedora pre-built is the shortest path to that setup.

Price it against the cloud. At $6,799, the machine only pays off if you run large models often, need your data to stay on your own hardware, or both. Compare it with a year of renting cloud GPUs for the same workload.

Move early if you want one this year. Framework is building a single first batch from the memory it has on hand, so later orders may wait longer than November.