AMD surfaces on‑device AI references

Artificial intelligence can run in the cloud or on a personal computer, and Advanced Micro Devices is now pushing the second option with OpenClaw, RyzenClaw, and RadeonClaw for local AI agents. (techspot.com) OpenClaw is the software layer, and AMD described RyzenClaw and RadeonClaw in mid-March 2026 as two hardware reference paths for running large language models on Windows machines instead of remote servers. The setup uses Windows Subsystem for Linux 2, with LM Studio handling local inference through the llama.cpp backend. (techpowerup.com) RyzenClaw is built around a Ryzen AI Max+ system with 128 gigabytes of unified memory, and AMD recommends reserving 96 gigabytes of that pool for graphics use in this workload. On that configuration, AMD said the Qwen 3.5 35B A3B model runs at about 45 tokens per second, handles a 10,000-token prompt in about 19.5 seconds, and supports a 260,000-token context window. (techspot.com) RadeonClaw shifts the work to a discrete Radeon AI PRO R9700 graphics card with 32 gigabytes of video memory. AMD said that path reaches about 120 tokens per second on the same model, cuts 10,000-token processing time to about 4.4 seconds, and supports a 190,000-token context window with up to two concurrent agents. (techpowerup.com) A token is a chunk of text a model reads or writes, and tokens per second is the speedometer for local inference. A context window is the working memory, and AMD’s pitch is that bigger local memory pools let a personal machine keep more documents and instructions in play without sending them to a data center. (techspot.com) AMD is framing that as part of what it calls an “Agent Computer,” a machine that can run always-on assistants with less network dependence and fewer outside services. TechPowerUp reported AMD said the environment can be configured in under an hour and is aimed at developers and early adopters rather than mass-market buyers. (techpowerup.com) The cost helps explain that positioning. TechPowerUp said a Framework Desktop with a Ryzen AI Max+ 395 and 128 gigabytes of memory starts at $2,700, while the Radeon AI PRO R9700 card starts at $1,299 before the rest of the system. (techpowerup.com) OpenClaw itself is not AMD-only software. The project’s site says it runs on macOS, Windows, and Linux, works with Anthropic, OpenAI, or local models, and includes browser control, file access, shell commands, and persistent memory features. (openclaws.io) An AMD-focused quick-start repository published in April 2026 shows how the company’s reference path is being translated into install scripts for Windows and Windows Subsystem for Linux 2. That guide points users to Ubuntu 24.04, LM Studio on the Windows host, and a one-command setup that connects OpenClaw to a local model endpoint with no application programming interface key. (github.com) The immediate shift is not that cloud AI is disappearing. It is that AMD is trying to make local inference concrete, with named hardware recipes and published performance targets, so developers can test how much agent work a desk-side machine can handle before reaching for the cloud. (techspot.com)