AMD and Unsloth announced on 19 November 2026 that Unsloth Desktop can now run local AI models on the NPU in AMD Ryzen AI processors. AMD says XDNA 2 NPUs are the first NPUs Unsloth Desktop supports. This article is based on AMD's announcement, AMD's blog and a report from Phoronix. Forge PC hasn't tried the NPU mode yet, so there are no speed figures here.
What's new
Unsloth is best known for its tools for fine-tuning language models, which we covered in how to fine-tune an LLM locally with Unsloth. According to AMD, Unsloth Desktop offers an OpenAI-compatible API endpoint that other apps can use.
With this update, those requests can be served by the NPU instead of the graphics card. AMD says the integration uses its open-source Lemonade server and FastFlowLM (FLM) for model serving and inference on the NPU. If you've read our piece on FastFlowLM adding GPU support, this is the same runtime doing the NPU work underneath. Lemonade is covered in how to run local AI with Lemonade Server.
100+ agent apps
AMD's headline claim is that the Unsloth Desktop API opens NPU inference to more than 100 compatible agentic AI applications and workflows. Its announcement names Hermes Agent, OpenClaw, Open WebUI, Cline, LangChain and GitHub Copilot as examples, and its blog also lists OMP and VS Code.
The idea, in AMD's words, is to use the NPU for "sustained, low-power inference" while leaving the CPU and GPU free for other applications. AMD says that can help preserve battery life on laptops, "memory permitting".
What you need
An AMD Ryzen AI PC with an XDNA 2 NPU. Phoronix says support starts with Ryzen AI NPUs of XDNA 2 and later. Not sure what your chip has? See what an NPU is.
Unsloth Desktop. AMD's steps use the Windows version. Phoronix says AMD confirmed to it that NPU support also works with the Linux builds.
Enough RAM. NPU models share system memory. AMD's example is a Qwen 3.6 35B A3B model, which it says needs around 28GB available to the NPU at full context. With Windows' default 50% shared allocation, AMD says that means 64GB of system RAM. Our guide to how much RAM you need for local AI covers this in more detail.
How to turn it on
AMD's blog gives these steps:
Download and install Unsloth Desktop for Windows on a PC with an XDNA 2 NPU.
Open the model hub page, click the filter icon and choose NPU from the drop-down list.
Click Enable NPU.
Download the NPU model you want. AMD notes that large downloads can look stuck when they aren't.
Go back to New Chat, scroll to the NPU section in the model drop-down, load the model with the context length you want and start chatting.
To confirm the NPU is actually doing the work, see how to check NPU usage in Windows 11.
Should you use the NPU or the GPU?
AMD describes XDNA 2 NPUs as designed for power-efficient AI inference, and its pitch here is about power and keeping other parts of the PC free, not about beating a graphics card. In our view, if you have a capable discrete GPU and run models while plugged in, it may still be the better choice. Our NPU vs GPU for local AI comparison explains the trade-offs.
Our verdict
Unsloth Desktop's NPU support gives Ryzen AI owners another practical use for a part of the chip that often sits idle. Any app that talks to Unsloth's OpenAI-compatible API can now use the XDNA 2 NPU, via Lemonade and FastFlowLM, on Windows and, per Phoronix, Linux. The catch is memory: AMD's own example of a 35B model needs a 64GB system, so check your RAM before downloading the bigger models.