Framework Desktop

What it is

Framework’s small desktop built around AMD’s Strix Halo chip. Unified memory up to 128GB, a fat iGPU, and the Framework promise: replace the parts instead of the whole box.

Who it is for

Builders who want a 24B–70B local model on the desk, also use the machine as a real PC, and do not want NVIDIA’s software leash. Good for llama.cpp / ROCm / Vulkan. Bad if your tools assume CUDA only.

What it actually runs

• Comfortable: 24B–32B dense at usable speed; larger MoE when you buy high-RAM configs.
• Possible: 70B-class if you accept that 256 GB/s is not Mac Studio bandwidth.
• Not the win: training or anything that wants a 5090 memory clock.

Standout details

• Repairable: replace the parts instead of the whole box.
• Unified memory up to 128GB. Bandwidth is ~256 GB/s — not Mac Studio.
• DIY assembly. Not CUDA; llama.cpp / ROCm / Vulkan.
• Street $1,269–$3,449 depending on config.

Honest verdict

The default recommendation if you want one machine for work and weights, and you can live without CUDA. If you want the same chip already assembled, see EVO-X2. If you only care about tokens and already own NVIDIA tooling, look at Spark / GX10 instead. Side-by-side: Framework vs EVO-X2.