AMD explains how to run OpenClaw on their hardware
AMD has published a new OpenClaw guide built around what it calls “ RyzenClaw ” and “ RadeonClaw ,” two AMD hardware paths for running local AI agents on Windows. The first is a Ryzen AI Max+ system with 128GB of unified memory, while the second uses a Radeon AI PRO R9700 graphics card. AMD is pitching both as ways to run OpenClaw locally through WSL2, LM Studio, and a local model setup rather than relying on the cloud.
Source: AMD
RyzenClaw & RadeonClaw
AMD’s “RyzenClaw” setup is not just any Ryzen AI Max+ laptop or mini PC, it specifically calls for a 128GB configuration, and AMD’s own instructions say users should reserve 96GB of variable graphics memory for that platform. That setup runs the Qwen 3.5 35B A3B model at about 45 tokens per second, supports a 260K token context window, and can handle up to six concurrent agents.
Source: AMD
The “RadeonClaw” path is faster, but it is not exactly consumer-friendly either. AMD pairs OpenClaw with its Radeon AI PRO R9700, a workstation card with 32GB of VRAM. AMD says that configuration pushes roughly 120 tokens per second with the same model and processes 10,000 input tokens in about 4.4 seconds, but it supports fewer concurrent agents than the 128GB Ryzen AI Max+ setup.
But who even has 128GB memory?
That was my first thought when AMD published this blog post. The cheapest US Strix Halo mini-PC with 128GB that we could verify is still somethng around $2,399, and that is before you even start talking about how niche this class of system already is. On the GPU side, AMD’s recommended Radeon AI PRO R9700 starts at $1,299.99 in the US, which is workstation pricing, not something the average gaming PC owner is likely to have sitting around.
Source: AMD
