AMD and Unsloth Bring NPU Acceleration to Desktop
The collaboration lets users run AI models locally on AMD's XDNA 2 neural processing unit through Unsloth Desktop.
Updated on Nov. 20, 2026 in Artificial Intelligence

Live Poll
Would you run AI models locally on your own computer instead of using online services?
AMD announced a collaboration with Unsloth to enable NPU acceleration in Unsloth Desktop, with AMD's XDNA 2 architecture as the first NPU architecture the software supports. AMD demonstrated the 35-billion-parameter Qwen3.6-35B-A3B model running on the NPU through the platform.
Why it matters
Moving supported AI workloads onto the NPU reduces CPU and GPU utilization, leaving those resources free for other applications. The approach lets locally hosted models run on Windows PCs without each application needing its own NPU inference engine.
AMD demonstrated Qwen3.6-35B-A3B, a 35-billion-parameter mixture-of-experts model, requiring about 28GB of NPU-accessible memory at full context, with Windows defaulting to 50% of shared system memory and AMD recommending 64GB of RAM. Independent benchmarks of throughput and power draw on this configuration have not been reported.
The players
AMD
Advanced Micro Devices is a global semiconductor company that designs CPUs, GPUs and AI accelerators, including the XDNA neural processing units in its Ryzen AI chips.
Unsloth
Unsloth is an open-source company that develops tools for running and fine-tuning large language models efficiently on local hardware.
The details
Unsloth Desktop uses Lemonade and FastFlowLM to execute inference directly on the NPU, and its OpenAI-compatible API lets applications such as Hermes Agent, OpenClaw, Open WebUI, Cline, OMP and Visual Studio Code reach locally hosted models. Users select supported models through the NPU filter in the model hub.
Timeline
• November 19, 2026: AMD announced the Unsloth NPU collaboration.
• July: Unsloth added support for Radeon, Ryzen AI and Instinct GPUs.
Tech Race
The integration places AMD's XDNA 2 NPU, the first architecture Unsloth Desktop supports, in direct competition with rival local-inference hardware for AI workloads on consumer PCs. It extends the NPU from a background accelerator into a platform that AI agent tools can call directly.
Windows users with AMD NPU hardware could run supported models locally, and moving those workloads to the NPU reduces CPU and GPU load for other tasks. Users with less than 64GB of RAM may find the full-context Qwen3.6 configuration hard to run, since AMD recommends that memory level.
The takeaway
Local AI tools that run on a PC's dedicated NPU can reduce load on the CPU and GPU, so checking whether your hardware and software support NPU acceleration can help you decide how to run AI models at home.
Further reading
For more on how local models are changing AI tools, see Artificial Intelligence.
Source note: This article includes information reported by VideoCardz.
Live Poll
Would you run AI models locally on your own computer instead of using online services?



