Using the latest Adrenalin Edition 25.8.1 driver, users with enough cooling and power can now load models like Meta’s Llama 4 Scout directly onto their desktop and run it without bowing to the whims of the cloud.

The trick lies in the company’s Ryzen AI MAX+ platform, specifically the Strix Halo APUs, which now support up to 96GB of shared graphics memory. That’s the sort of footprint usually demanded by enterprise-grade inferencing, now showing up in consumer gear you can plug into a regular ATX board.

XDNA engines do the AI legwork, but the real story is memory capacity. While competitors are still mucking about with half-baked AI widgets, AMD’s shoved a truckload of VRAM into consumer-facing systems and declared open season on edge inferencing.

You’ll still need a decent dGPU and plenty of juice, but the barrier to entry for local LLM deployment just got a lot lower. With this move, AMD has effectively told developers, researchers, and data hoarders they don’t need to rent someone else's silicon to get serious work done.