AMD and Unsloth are putting XDNA 2 neural processing units (NPUs) to work on local AI agent requests. In a November 19 announcement, AMD says more than 100 compatible agentic applications and workflows can access local model inference through Unsloth Desktop on supported Ryzen AI PCs. The GPU isn’t being handed a retirement cake. Source Source
AMD says XDNA 2 NPUs are the first NPUs supported by Unsloth Desktop. And Windows doesn’t have exclusive custody: Phoronix’s Michael Larabel reports that AMD confirmed the support also works with Linux builds. Source Source
What actually moves onto the NPU
Agents can carry out multistep tasks across connected tools. The NPU’s job here is supported model inference: processing a model request and producing a response. According to AMD, a compatible application sends its request to the Unsloth Desktop API, while Lemonade and FastFlowLM handle model serving and inference on the XDNA 2 NPU. The response returns to the application, whose other functions continue to use the CPU and GPU as needed. Source
Unsloth Desktop offers an OpenAI-compliant API endpoint for compatible tools. AMD names Hermes Agent, OpenClaw, Open WebUI, Cline, LangChain and GitHub Copilot among the ecosystem examples. The 100-plus figure covers applications and workflows together—not necessarily more than 100 distinct apps. Source
AMD says offloading supported inference can help preserve battery life and leave CPU and GPU resources available for other applications, memory permitting. Its announcement and setup post don’t provide measured battery-life, power or speed comparisons. The efficiency pitch is there; the benchmark receipts aren’t. Source Source
Getting started—and checking the RAM bill
AMD’s Windows guide says to install Unsloth Desktop on a device with an XDNA 2 NPU, filter the model hub for NPU models, select Enable NPU, and download a supported model. Once downloaded, it can be loaded from the NPU section in New Chat with the desired context. The guide doesn’t cover Linux setup. Source
AMD says Qwen 3.6 35B A3B at full context needs around 28GB of memory available to the NPU. Under the Windows default 50% shared allocation described in its guide, the company says that requires 64GB of system RAM. That’s a model-specific example, not a universal minimum for Unsloth Desktop or a stated Linux requirement. Source
Comments