AMD and Unsloth have brought NPU support to Unsloth Desktop, so people with recent Ryzen AI laptops can run local language models on the chip’s built-in AI engine instead of the graphics card. Through Unsloth’s OpenAI-compatible API, the setup can serve more than 100 agentic apps and workflows, including Hermes Agent, OpenClaw, Open WebUI, Cline and GitHub Copilot, AMD said on 19 November 2026.
Quick Take
- What’s new: AMD XDNA 2 NPUs are the first NPUs supported by Unsloth Desktop, the local app for running and training models.
- Hardware: Ryzen AI 300 (Strix Point, Krackan Point), Ryzen AI Max (Strix Halo) and Ryzen AI 400 (Gorgon Point), on Windows 11 or x64 Linux.
- Software path: apps call the Unsloth Desktop API; Lemonade and FastFlowLM serve the model on the NPU.
- Biggest model shown: Qwen3.6 35B A3B, a mixture-of-experts model, runs at 17.2 tokens per second in Unsloth’s measurements and needs a 64 GB system.
- Efficiency claim: Unsloth says NPUs use “2 to 4x less energy per token generation than traditional GPUs.”
In plain English
Many new laptops have a special chip for AI, called an NPU, that mostly sits idle. AMD and Unsloth now let that chip run a chatbot or coding assistant locally, without sending data to the cloud and without tying up the graphics card. Apps that already talk to OpenAI-style services can point at the laptop instead. The trade-off is speed: larger models run noticeably slower than on a strong GPU.
What happened
AMD announced the collaboration in a newsroom post and a technical blog by Syed Muhammad Usman Pirzada. “With support for Unsloth Desktop, AMD is expanding the supported compute resources for developers and users to deploy and run agentic workflows locally, efficiently and through familiar applications,” the company said.
Unsloth published a setup guide the same day. Users open the model picker, switch the filter to NPU and click Enable NPU; Unsloth then downloads FastFlowLM and checks the driver version.
Background
Unsloth is best known for open-source tools that speed up fine-tuning and cut GPU memory needs, and we covered its Unsloth Studio app earlier. Local agents are a fast-growing niche: projects like Linger run a large model on a 12 GB gaming GPU. AMD has been pushing its Ryzen AI line as a platform for on-device AI alongside its data center GPUs.
How it works
- Model format: the NPU runs 4-bit Q4NX models built for AMD’s FastFlowLM runtime.
- Measured speeds (Unsloth): Qwen3.6-35B-A3B at 17.2 tokens per second with 3.13 seconds to first token; Gemma4-12B-IT at 6.8 tokens per second; Qwen3.5-9B at 9.5; Qwen3-VL-4B at 20.5; LFM3-1.7B at 40.7.
- Memory: AMD notes the 35B model with full context needs about 28 GB available to the NPU, which with Windows’ default 50% shared allocation means 64 GB of system RAM.
- Features: streaming, reasoning, vision, tools, MCP servers, web search and chatting with local files.
- Linux: requires kernel 7.0 or newer with the amdxdna driver, or the amdxdna-dkms package, plus updated NPU firmware.
- Power modes: power saver, balanced, performance or turbo.
Why it matters
For laptop users, it puts an often unused part of the processor to work, which can save battery and keep the GPU free for games, video or rendering.
For developers, an OpenAI-compatible local endpoint means coding tools and agent frameworks can switch to on-device inference with a configuration change, which helps with privacy and offline work.
For the PC market, NPUs have so far lacked compelling software. Support in a popular local-model app is the kind of integration that gives AI PC branding practical meaning.
What’s still open
- Speeds for larger models are modest; the 27B Qwen3.8 model is listed at 2.0 tokens per second.
- The energy-efficiency figure is Unsloth’s claim, without a published comparison setup.
- Whether Intel and Qualcomm NPUs get similar support in Unsloth.
Sources
- AMD Newsroom, “AMD and Unsloth Bring NPU Support to 100+ Agentic AI Applications,” 19 November 2026
- AMD blog, “AMD x Unsloth: Unlocking the World’s First NPU for Unsloth Desktop,” 19 November 2026
- Unsloth documentation, “AMD NPU: Run Models Locally Guide”
i10x Perspective
Local AI is moving from enthusiasts’ GPUs to the chips already inside mainstream laptops. Slower, yes, but private, cheap and always on, which suits background agents well. More on AI hardware in Infrastructure.



