Official AMD and Unsloth illustration of XDNA 2 NPU support
Featured

AMD and Unsloth bring local NPU support to more than 100 AI workflows

AMD and Unsloth announced on November 19, 2026 that compatible Ryzen AI PCs can use the AMD XDNA 2 neural processing unit for more than 100 agentic AI applications and workflows through Unsloth Desktop. XDNA 2 becomes the first NPU architecture supported by the desktop software, extending local AI beyond workloads that rely entirely on a CPU or GPU.

The change is primarily about distributing work across the computer. Model inference can run on the NPU while the CPU and GPU remain available for other tasks. That could matter when an assistant stays active alongside video editing, games or everyday applications, although available memory and the particular model still determine what a system can handle.

How does support for more than 100 applications work?

Unsloth Desktop exposes an OpenAI-compatible API endpoint. A compatible application sends its request to that endpoint; the Lemonade and FastFlowLM runtime stack handles supported local inference on the NPU and returns the result. API compatibility does not mean that OpenAI's proprietary models are downloaded to the computer, nor that every component of an application moves onto the NPU.

AMD names examples including Hermes Agent, OpenClaw, OpenWebUI, Cline, LangChain and GitHub Copilot. The headline figure describes compatible applications and workflows, rather than 100 separately benchmarked performance gains or a promise that every model used by those applications is supported. Developers still need to select an appropriate local model and configure the endpoint.

Which computers and models are supported?

Unsloth's documentation lists Ryzen AI 300, Ryzen AI Max and Ryzen AI 400 systems among the supported hardware families, with XDNA 2 or newer NPU hardware. Windows 11 and Linux x64 are covered. Linux also requires suitable drivers, firmware and memory-lock settings. An AMD badge alone is therefore not enough to establish compatibility: the processor generation and software environment matter.

The NPU path uses compatible FastFlowLM models in the four-bit Q4NX format. Users can filter the model catalogue for NPU support, enable the NPU option to obtain the runtime and check drivers, download a compatible model, then select it in a new chat and configure the context. This is not a universal switch that makes every existing model run on the NPU.

Official AMD and Unsloth illustration of XDNA 2 NPU support
AMD's official announcement artwork, fitted into a 4:3 frame while preserving its content.

Why does memory capacity matter?

AMD illustrates the memory requirements with Qwen3.6-35B-A3B, a mixture-of-experts model with 35 billion total parameters. Its full-context configuration needs roughly 28 GB of NPU-accessible memory. With Windows allocating up to half of system memory to the shared pool in this example, AMD points to a 64 GB machine. That is a requirement for the illustrated configuration, not a minimum for every supported model or chat.

Unsloth can size context automatically according to available memory. Larger models, longer conversation context and image inputs can change the resource demands. Power saver, balanced, performance and turbo modes also offer different operating choices, so users should match the configuration to the task rather than simply selecting the largest model.

A battery-life objective, not a universal benchmark guarantee

AMD positions the NPU as a way to keep persistent agent workloads running with lower power consumption while freeing CPU and GPU resources. The announcement does not establish one battery-life improvement that applies to every laptop. Actual results depend on the model, context size, power mode, memory capacity and the rest of the system's workload.

Unsloth is also known for open-source fine-tuning tools, and its desktop software includes training and inference features. This announcement concerns NPU inference; it should not be read as moving all model training, image generation or video generation onto the NPU. Likewise, running a model locally does not automatically make an agent entirely offline. Connected tools and services may still send data elsewhere, so their permissions and destinations remain important.

The practical significance is a more accessible route from an NPU inside a computer to everyday AI software. A no-code desktop interface and a familiar API give users and developers another local execution option. The value of that option will depend on compatible hardware, suitable models and realistic expectations about memory and performance.

Sources: AMD announcement, AMD technical blog, Unsloth NPU documentation.

Image

Subscribe for Updates