Framework Desktop (Ryzen AI Max+ 395)
No editorial image yet — generic vendor mark shown. Credentials in spec table below.
The enthusiast-favorite Strix Halo box: a Ryzen AI Max+ 395 system with 128GB LPDDR5X-8000 unified memory (~256 GB/s), up to ~96GB allocatable as VRAM on Linux. ~$1,999 DIY, open hardware, Linux-first. A direct DGX Spark / Mac Studio competitor that runs 70B+ models with no discrete GPU.
Affiliate disclosure: as an Amazon Associate and partner of other retailers, we earn from qualifying purchases. The verdict on this page is our editorial opinion; affiliate links never influence what we recommend.
Sub-scores sum to 219 / 1000. Headline = 219 × 0.70 (Estimated-confidence discount) = 153. This is an algorithmic performance-tier score — distinct from, and often lower than, the editorial “Our verdict” below, which weighs value and real-world fit (especially for hardware we haven’t measured yet). How scoring works →
Extrapolated from 256 GB/s bandwidth — 25.6 tok/s estimated. No measured benchmarks yet.
Plain-English: Doesn't fit modern chat models usefully — vision models won't fit.
Verdicts extrapolated from catalog VRAM + bandwidth + ecosystem flags. Hover any chip for the rationale. Want measured numbers? Submit your own run with runlocalai-bench --submit.
What it does well
Framework Desktop is the open-hardware answer to the Mac Studio for local AI. The Ryzen AI Max+ 395 (Strix Halo) exposes up to ~96GB of its 128GB unified pool as VRAM on Linux, so it runs Llama 3.3 70B at Q6 conversationally and fits models no consumer discrete GPU can — at ~140W in a small repairable box for ~$1,999. It's the pick for tinkerers who want x86 + Linux + ROCm/Vulkan rather than Apple's walled garden, and Framework explicitly supports multi-board LLM clustering.
Where it struggles
256 GB/s unified bandwidth is roughly half a Mac Studio M4 Max and a fraction of a discrete GPU, so token-generation on big models is slower than the VRAM capacity implies — this is a 'fits huge models' machine more than a 'fast' one. ROCm on Strix Halo is improving fast but still has rough edges and is Linux-only for serious GPU offload; expect some setup work versus an NVIDIA box. No CUDA.
Bottom line
The best open/Linux-first way to fit 70B-class models locally for ~$2k. Buy it over a Mac if you want repairability, x86, and Linux; accept that bandwidth-bound token speed trails both NVIDIA and the higher-bandwidth Macs.
Overview
The enthusiast-favorite Strix Halo box: a Ryzen AI Max+ 395 system with 128GB LPDDR5X-8000 unified memory (~256 GB/s), up to ~96GB allocatable as VRAM on Linux. ~$1,999 DIY, open hardware, Linux-first. A direct DGX Spark / Mac Studio competitor that runs 70B+ models with no discrete GPU.
Some links above are affiliate links. We may earn a commission at no extra cost to you. How we make money.
Specs
| System RAM (typical) | 128 GB |
| Power draw (peak) | 140 W |
| Released | 2025 |
| MSRP | $1999 |
| Backends | ROCm Vulkan |
Models that fit
Open-weight models small enough to run on Framework Desktop (Ryzen AI Max+ 395) with usable context.
Frequently asked
Does Framework Desktop (Ryzen AI Max+ 395) support CUDA?
Where next?
Reviewed by RunLocalAI Editorial. See our editorial policy for how we research and verify hardware specifications.