News

Unsloth Desktop adds AMD XDNA 2 NPU inference on Ryzen AI PCs

Unsloth Desktop now runs supported local models on AMD XDNA 2 NPUs through Lemonade and FastFlowLM, opening a dedicated inference path for compatible Ryzen AI PCs on Windows and Linux.

D
Nov 19, 2026 · 1 min read

AMD and Unsloth have added local NPU inference to Unsloth Desktop, starting with AMD XDNA 2 hardware. The new path gives compatible Ryzen AI PCs a dedicated processor for supported local models on Windows 11 and Linux.

Unsloth’s setup guide lists Ryzen AI 300 systems based on Strix Point or Krackan Point, Ryzen AI Max systems based on Strix Halo, and Ryzen AI 400 systems based on Gorgon Point. Each requires an XDNA 2 or newer NPU. Designed for AI workloads, the NPU sits alongside the system’s general-purpose CPU and graphics processor.

Compatible applications send model requests to Unsloth Desktop through an OpenAI-compatible API. Lemonade and FastFlowLM then handle model serving and inference on the NPU before Unsloth Desktop returns the response. The NPU path uses Q4NX 4-bit models built for FastFlowLM. Arbitrary GGUF models cannot run on it directly, although Unsloth says supported GGUF models can be converted to Q4NX.

AMD says the API works with more than 100 agentic applications and workflows, including Hermes Agent, OpenClaw, Open WebUI, Cline, LangChain and GitHub Copilot. AMD also says moving supported inference to the XDNA 2 NPU can help preserve battery life and support persistent low-power agents while leaving CPU and GPU resources available. Unsloth claims the NPU uses two to four times less energy per generated token than traditional GPUs. The opened sources did not independently test the energy or battery-life claims or provide a complete inventory supporting AMD’s application count.

On Linux, Unsloth requires x64, kernel 7.0 or newer with the amdxdna driver, or an older kernel with amdxdna-dkms. Its guide also calls for updated NPU firmware and an unlimited memlock setting. Enabling NPU support in Unsloth Desktop downloads FastFlowLM and checks the installed NPU driver before presenting compatible models.

More news