AMD unveils OpenClaw to run AI agents locally on Ryzen and Radeon hardware

RyzenClaw and RadeonClaw aim to run LLMs and multi-agent workflows on local PCs

by · TechSpot

Serving tech enthusiasts for over 25 years.
TechSpot means tech analysis and advice you can trust.

The takeaway: AMD is pushing the idea that artificial intelligence agents don't need to live in the cloud. Its new OpenClaw framework – now equipped with two hardware configurations dubbed RyzenClaw and RadeonClaw – is designed to help developers and early adopters run sophisticated large language models entirely on local machines. The aim is clear: bring generative AI performance into the home and reduce dependence on data centers.

The effort is part of AMD's broader Agent Computer initiative, which argues that the future of AI isn't limited to remote infrastructure. Instead, it envisions a world where users control both their data and their computing environment – where AI assistants operate continuously with reduced network dependence, fewer external subscriptions, and fewer privacy concerns.

OpenClaw is AMD's latest attempt to turn that principle into a tangible, developer-accessible platform. At a technical level, OpenClaw runs on Windows using the Windows Subsystem for Linux (WSL2), with local inference handled by LM Studio through the llama.cpp backend. This setup allows users to run models such as Qwen 3.5 35B A3B directly on their own hardware.

The system also supports Memory.md, an embedding-based memory framework that stores local context without relying on cloud synchronization. AMD presents the guide as a streamlined way for developers to configure a full OpenClaw environment on Windows when testing AI agent architectures, though it does not specify an expected setup time.

The two configurations represent different paths to the same idea: high-performance, on-device AI. The RyzenClaw configuration is built around AMD's Ryzen AI Max+ processor paired with 128GB of unified memory. AMD recommends allocating roughly 96GB of that memory to variable graphics usage to keep LLM inference running efficiently.

In this configuration, Qwen 3.5 35B A3B generates about 45 tokens per second and can process a 10,000-token input in approximately 19.5 seconds. Its 260,000-token context window is expansive, making it suitable for multi-agent workflows or "agent swarm" testing environments. AMD says the setup can run up to six local AI agents concurrently – a notable figure for a non-datacenter system.

RadeonClaw, by contrast, shifts the computing load to a discrete GPU: the Radeon AI PRO R9700. This workstation card comes with 32GB of dedicated VRAM, which significantly increases throughput. Using the same model, performance climbs to around 120 tokens per second, reducing the time needed to process 10,000 tokens to about 4.4 seconds.

That gain, however, comes with limits as the maximum context window drops to 190,000 tokens, and concurrent agent capacity falls to two. These trade-offs underscore AMD's strategy of offering distinct tuning paths depending on whether developers prioritize context depth or inference speed.

// Related Stories

Neither configuration is built for casual users. A desktop built around the Ryzen AI Max+ 395 chip and 128GB of memory such as a Framework Desktop configuration is cited as starting at around $2,700. The RadeonClaw option adds further expense, as the Radeon AI PRO R9700 GPU alone retails for about $1,299.

For now, AMD acknowledges that OpenClaw targets early adopters and engineers experimenting with local AI agents rather than mainstream consumers.

Still, the message behind OpenClaw extends beyond its hardware. AMD is betting that developers will value autonomy and privacy over raw scale, and that local agents running on consumer-grade silicon can bridge the gap between personal computing and distributed AI.

If that idea gains traction, the company could carve out a distinct role in the rapidly evolving AI ecosystem – one that blurs the line between workstation and datacenter performance.