What we checked
No unit sat on a bench for this review. The check is the pages that were open. The date on this file is that day.
Framework's spec table for the Desktop lists the Ryzen AI Max+ 395 with 16 cores, 32 threads, a 3.0 GHz base, a 5.1 GHz boost, 64 MB of L3, and a Radeon 8060S with 40 compute units at up to 2.9 GHz. Memory on that config is 128 GB of LPDDR5x-8000, soldered, on a 256-bit bus at 8000 MT/s. Framework rates the processor at 120 W sustained and 140 W boost, in a 4.5 L case that weighs 3.1 kg, with a 400 W supply. The marketing page says you can run a Llama 70B with up to 96 GB of graphics-addressable memory.
AMD's product page for the Ryzen AI Max+ 395 agrees on the chip: 16 cores, 32 threads, up to 5.1 GHz, 64 MB of L3, Radeon 8060S, 40 compute units, 256-bit LPDDR5x, 128 GB max, LPDDR5x-8000. AMD lists a 55 W default TDP and a 45 W to 120 W configurable TDP. Framework's 140 W boost sits above that configurable ceiling. Both numbers are on the pages. This review does not pick a winner between them.
AMD's variable-graphics-memory FAQ, from 2025, is where the bandwidth sentence lives. In the paragraph about the 64 GB Strix Halo setup, AMD says the platform doubles bandwidth to 256 GB/s. The 128 GB paragraph does not restate that figure. It does say the 128 GB processor can dedicate 96 GB through variable graphics memory, run 4-bit models up to 128B parameters, run FP16 models up to 32B parameters, and reach 112 GB of total graphics-addressable memory with that 96 GB setting. Framework's spec table gives the 64 GB and 128 GB configs the same bus width and the same 8000 MT/s. Bandwidth was not measured here.
Prices, from Framework's own store on the same pass. The spec page says desktops start at $1,269. That is the floor, not this config. The 128 GB mainboard listed at $3,149 and was out of stock. The refurbished Desktop DIY edition listed the 128 GB config at $2,929 and was out of stock. That refurbished page says storage, the fan, and the OS are not included. A current new-condition Desktop price for the 128 GB config was not on the pages opened, so this review does not print one.
What you'll feel
The feel below is what those pages imply. It is not a session on the hardware.
You would feel the pool, not a speed record. 128 GB is why a 70B-class model stops being a quantization problem, which is the same reason the DGX Spark review exists. You would not feel a 5090. AMD's published Strix Halo bandwidth is 256 GB/s, and every decode-bound workload already on this site treats that class of pipe as the ceiling.
You would also feel the software. This is an AMD iGPU with a setting that turns system memory into a dedicated graphics block. It is not a CUDA card. Image and video stacks that assume NVIDIA stay on the 5090 review.
Setup notes
Memory is soldered. Order the 128 GB config or the argument does not apply. Variable graphics memory, on AMD's FAQ, is an Adrenalin toggle on Windows and it reboots. It subtracts from the RAM the CPU can see. AMD calls the enthusiast preset a high custom setting, and the FAQ says a bad setting can hurt the rest of the machine.
Framework's Desktop is a DIY chassis: a Mini-ITX board, a FlexATX supply, and a 120 mm fan you pick or bring. The spec page says bring your own OS, including Linux. AMD's product page lists Ubuntu and RHEL next to Windows. Neither was installed for this review.
The 140 W boost Framework prints and the 120 W configurable TDP AMD prints are not the same number. Plan around the 400 W supply Framework specs for the chassis. Do not treat 140 W as an AMD specification.
Who should buy
- A practitioner who needs a 128 GB pool on a desk and can work on AMD's graphics stack.
- Someone who wants a repairable Mini-ITX box rather than a sealed appliance, and who has already accepted soldered memory.
- A team prototyping a 70B-class model at a quant that fits in the 96 GB graphics block AMD describes.
Who should skip
- Anyone whose model already fits on a 5090 and who cares about decode speed. Bandwidth is why this is not that card.
- Anyone whose tools require CUDA.
- Anyone who read $1,269 on the spec page and thought that was the 128 GB machine.
Bottom line
Buy it if the job is capacity and the stack is AMD. Do not buy it for tokens per second, and do not buy it off the starting price. Scored the way the DGX Spark review is scored: the pool is real, the pipe is the limit, so the verdict is buy-if and the score is 7.