AMD has introduced OpenClaw, a new framework designed to run artificial intelligence agents directly on local machines rather than relying on cloud infrastructure. The platform includes two hardware configurations—RyzenClaw and RadeonClaw—built to deliver high-performance AI processing on personal systems powered by Ryzen and Radeon hardware.
The launch reflects AMD’s broader vision of enabling on-device generative AI, giving users greater control over their data while reducing dependence on remote data centers.
Part of AMD’s Agent Computer Initiative
The OpenClaw framework is part of AMD’s Agent Computer initiative, which promotes a future where AI systems operate locally on user devices rather than entirely in the cloud.
The approach aims to provide several benefits, including:
- Greater data privacy
- Reduced reliance on internet connectivity
- Lower subscription costs for AI services
- Faster AI responses through local processing
By enabling powerful local inference, AMD hopes to make personal AI assistants and multi-agent systems practical for developers and advanced users.
How OpenClaw Works
OpenClaw runs on Microsoft Windows using the Windows Subsystem for Linux (WSL2). Local AI inference is handled through LM Studio, powered by the llama.cpp backend.
This setup enables developers to run advanced large language models locally, including Qwen 3.5 35B A3B, directly on their own hardware without cloud-based processing.
The system also integrates Memory.md, an embedding-based memory framework that stores contextual information locally, eliminating the need for cloud synchronization.
RyzenClaw Configuration: High Context Capacity
The RyzenClaw configuration is designed for large context windows and multi-agent workflows. It runs on the AMD Ryzen AI Max+ processor paired with 128GB of unified memory.
AMD recommends allocating approximately 96GB of memory to variable graphics usage to maintain efficient large language model inference.
Performance highlights include:
- Around 45 tokens per second generation speed
- Processing 10,000 tokens in about 19.5 seconds
- A large 260,000-token context window
- Ability to run up to six local AI agents simultaneously
This setup is particularly suitable for multi-agent systems or experimental “agent swarm” environments.
RadeonClaw Configuration: Higher Speed with GPU Acceleration
The RadeonClaw configuration shifts the workload to a discrete GPU, specifically the Radeon AI PRO R9700.
This workstation GPU includes 32GB of dedicated VRAM, significantly improving inference speed.
Key performance metrics include:
- About 120 tokens per second generation speed
- Processing 10,000 tokens in roughly 4.4 seconds
However, this configuration involves trade-offs. The maximum context window drops to 190,000 tokens, and the system can run two AI agents concurrently instead of six.
A Step Toward Local AI Ecosystems
With OpenClaw, AMD is emphasizing a future where powerful AI capabilities can run directly on personal hardware rather than relying entirely on cloud platforms.
By offering different configurations optimized for either deeper context handling or faster inference speeds, the company aims to give developers flexibility when building next-generation AI agents and local AI applications.
