Perplexity Brings Local Agent Runtime to AMD Ryzen AI Max PCs

Perplexity's Portable Computer is the local version of its Computer agent: the agent harness, orchestrator, sandbox and a 27-billion-parameter model all run on the user's own PC instead of in Perplexity's cloud. Its Windows version first shipped for PCs with Nvidia RTX graphics, as Nvidia described in September. This announcement adds Windows machines built on AMD's Ryzen AI Max Series processors.

AMD's announcement gives the reason these chips qualify: Ryzen AI Max processors combine CPU and GPU performance with a large pool of shared memory, which AMD says lets them run local models alongside everyday applications. Perplexity's product page sets the bar at 24GB of VRAM or more, the same threshold it uses for RTX cards, and lists AMD support for Windows only; its Linux instructions cover Nvidia hardware. On a Windows PC, the page has users set up Perplexity's own PPLX 27B for local inference (it lists Qwen 3.8 27B as a second option on Nvidia's DGX Spark), and only one local model runs at a time. The product is available to Pro and Max subscribers.

The post says scheduled work runs "entirely on your device", and Perplexity's product page describes that as the default rather than a hard boundary. When a step needs the web or stronger reasoning than the local model offers, the orchestrator asks for approval before routing that step to one of more than 15 cloud models, then returns the result to the local run. Connected apps such as Gmail, Slack and GitHub also go through the local orchestrator. AMD and Perplexity both point to cost: work done by the local model does not use Perplexity Computer credits and carries no per-token charge, which Perplexity says makes large, repetitive jobs practical on hardware the user already owns.