Framework Desktop (Ryzen AI Max+ 395) review

If you’re pricing a 128GB box to run models at home, the Framework Desktop will be on your list, and it’s built from parts you’d recognise from any desktop PC. It uses a standard Mini-ITX board, a FlexATX power supply and a 120mm fan mount. The board carries AMD’s Ryzen AI Max+ 395 (the chip codenamed Strix Halo) with 128GB of LPDDR5X soldered on, which means the memory can’t be upgraded later.

The take

You can’t order the 128GB Ryzen AI Max+ 395 system, which is out of stock, but Framework lists it at $3,449 at the time of writing (12 November 2026). That’s the DIY build, which comes without an SSD, an operating system or a fan. The smaller configurations are out of stock as well, a 32GB at $1,269 and a 64GB at $1,959.

In its November 2025 review, ServeTheHome said that a few weeks earlier, many would have called $1,999 a high base price for a 128GB system. By Framework’s own account in January 2026, the price of the memory chips in its 128GB build had spiked the most.

Even so, at the time of writing the 128GB Framework is about $50 under the GMKtec EVO-X2 at about $3,500, although the GMKtec’s price includes a 1TB SSD. The Beelink GTR9 Pro is $4,199 on pre-order. The DGX Spark is $6,950, so the Framework is half the money. A 128GB Mac Studio with the M5 Max is $5,099 with a 512GB SSD or $5,399 with 1TB, which puts the Framework at 64 to 68% of it. The Spark’s memory is a little quicker on paper, 273 GB/s against the Framework’s 256 GB/s.

Framework now leads its desktop range with a 192GB model on the Ryzen AI Max+ PRO 495. It’s a 16-core, 32-thread chip with 192GB of LPDDR5X-8533 on the same 256-bit bus, and an NPU rated at up to 55 TOPS. That memory speed works out at about 273 GB/s, which matches the Spark. The 495’s PCIe x4 slot is open-ended and supplies up to 25W, though the default case doesn’t expose it. The DIY version is $6,799 on pre-order at the time of writing, with the November batches sold out and Batch 3 shipping in December. The pre-built starts at $7,449.

What it’ll run

A model on the Framework isn’t held to the VRAM on a graphics card, because the memory is unified (one pool that the processor and the graphics core share). AMD says the graphics core can reach 112GB of the 128GB. Framework’s own figure is 96GB of VRAM in Windows, with Linux able to go higher, although Jeff Geerling hit errors above about 108GB.

A 70B model at Q4 is 42.5GB, less than half of the 96GB Windows allows. The model Framework suggests for the 128GB box, DeepSeek V4 Flash 0731 at UD-IQ2_XXS, is 90.9GB. On pre-production units in August 2025, Geerling had gpt-oss-120b at about 33 tok/s, gpt-oss-20b at about 45 and Qwen2.5 14B at about 23.

On a dense model, the Framework’s token speed comes down to memory bandwidth (how fast the chip can read its memory), since every token means reading all of the weights. Strix Halo’s memory bandwidth is 256 GB/s in theory, and lhl on the Level1Techs forum measured about 215 GB/s. Divide that by the 42.5GB of a 70B at Q4 and the ceiling is about 5 tok/s, which is almost exactly the 4.97 Geerling measured. Double figures at that size takes a mixture-of-experts (MoE) model like gpt-oss-120b, which only works through part of its weights for each token. At the other end, a 3B ran at 88 tok/s in those tests.

The 50 TOPS NPU doesn’t help llama.cpp, because AMD’s route to it is ONNX Runtime GenAI. Vulkan was ahead of ROCm in Geerling’s August 2025 tests, but kyuz0’s Strix Halo toolboxes now recommend ROCm for performance. In the same tests, Ollama seemed to use only the CPU cores on its defaults, and an owner reported in a GitHub issue that Ollama’s ROCm provider saw just the 512MB set aside in the BIOS.

Who should buy it

The Framework makes sense if you want 128GB for less than the GMKtec, the Beelink, the Spark or a Mac Studio, going by list prices at the time of writing, and you can wait for stock. You’ll need to budget on top for an SSD, a fan and an operating system, which can be Windows or Linux. If you’d rather fix a machine than replace it, the board, power supply and fan are all standard sizes, and Framework sells spares through its Marketplace. There are two M.2 2280 sockets that take up to 8TB each. The warranty is a year, with three years as a paid option.

It’s the wrong machine if you want fast answers from a big dense model, with a 70B at Q4 managing only about 5 tok/s. And there’s no CUDA on AMD graphics, which rules it out if your tools depend on it. The 128GB build is still out of stock, even though the configurator says it ships within 5 business days. The UK store also warns that deliveries to Europe are delayed by up to 7 days.