News
AMD and Unsloth Enable World's First NPU Support for Unsloth Desktop
AMD XDNA™ 2 NPUs now power local AI inference on Unsloth Desktop, unlocking 100+ agentic apps with efficient, low-power execution.
AMD and Unsloth have announced a groundbreaking collaboration to bring neural processing unit (NPU) support to Unsloth Desktop, marking the first time NPUs are integrated into the application. This development leverages AMD's XDNA™ 2 architecture, enabling local AI inference for over 100 agentic applications and workflows.
Expanding Local AI Capabilities
The integration is powered by AMD's open-source tools, Lemonade and FastFlowLM (FLM), which facilitate model serving and inference on the AMD XDNA™ 2 NPU. Through the Unsloth Desktop API, which offers an OpenAI-compliant endpoint, users can connect applications such as Hermes Agent, OpenClaw, and VS Code to local AI models. This ecosystem enables efficient, low-power execution of AI agents directly on PCs, reducing reliance on GPUs and CPUs while preserving battery life.
Technical Highlights
To utilize this feature, users must install Unsloth Desktop on an AMD device equipped with an XDNA™ 2 NPU. Supported models, such as the Qwen 3.6 35B A3B, require specific memory allocations—28GB of NPU memory for full context, translating to a system with at least 64GB of RAM under Windows' default shared allocation settings. Once configured, users can run local inference seamlessly through the NPU, as demonstrated by a 35-billion-parameter mixture-of-experts model in real-time testing.
Implications for Developers and Users
This collaboration opens new possibilities for developers to integrate low-power, persistent AI agents into everyday workflows. For users, it provides a more efficient way to run AI applications, particularly in unplugged scenarios where battery life is critical. By offloading inference workloads to the NPU, AMD and Unsloth are redefining how local AI can be deployed and experienced on PCs.
With this innovation, AMD and Unsloth are setting a new standard for AI acceleration, making advanced agentic applications more accessible and efficient for a broad range of users and developers.