The AI chip you already paid for
If you bought a Windows laptop in the last couple of years, there is a good chance it has a part you have never knowingly used. The sticker said Copilot+ or Ryzen AI. The spec sheet called it an NPU, a neural processing unit. Microsoft requires one rated at 40 or more trillion operations a second for a laptop to be sold as a Copilot+ PC, and AMD rates the Ryzen AI 300 series NPU at up to 50.
So far, Windows has used it for a short list of built-in tricks: Microsoft lists Recall, Cocreator, Windows Studio Effects, Live Captions and automatic super resolution. Running a chatbot you choose on it has meant extra runtimes and the command line. This week that got much easier, and this guide walks you through it.
Why this matters now
On Nov. 19, AMD and Unsloth announced that Unsloth Desktop, a free app for running and training AI models at home, can now run models on AMD's XDNA 2 NPUs. AMD calls them the first NPUs the app supports. The same day, Unsloth shipped release v0.1.920-beta, and Phoronix reported that AMD confirmed it works on Linux, not just Windows.
- It leaves your graphics chip free. Unsloth's own example is chatting with a model on the NPU while your Radeon GPU renders a video.
- It's built for battery. Unsloth's release notes say NPUs are 2 to 4 times more efficient per joule per token than GPUs. That is the vendor's figure; we found no independent test yet.
- Other apps can use it. Unsloth exposes an OpenAI-compatible endpoint, and AMD says more than 100 agent apps and workflows work with it, naming Hermes Agent, OpenClaw, Open WebUI, Cline, LangChain and GitHub Copilot.
- Nothing leaves the machine. Once the model is downloaded, your prompts and files stay on your laptop. No meter running, no data center.
What you need
Find your processor's name in Settings > System > About, or in Task Manager under Performance > CPU. Then find your row:
| Your processor | NPU | How to use it | Status, Nov. 20 |
|---|---|---|---|
| AMD Ryzen AI 300, Ryzen AI 400, Ryzen AI Max | XDNA 2 | Unsloth Desktop (Windows 11 or x64 Linux) | Supported since Nov. 19 |
| Older Ryzen AI chips (Ryzen 7000 and 8000 series laptops, Ryzen AI 200) | XDNA 1 | – | Not supported by FastFlowLM or Unsloth |
| Intel Core Ultra (Arrow Lake and later, per Microsoft) | Intel NPU | Microsoft Foundry Local (Windows) | Models that have an NPU build |
| Snapdragon X Elite, X Plus | Hexagon | Microsoft Foundry Local (Windows on Arm) | Models that have an NPU build |
| Apple Silicon Mac | Neural Engine | None of these tools | Use the GPU instead (see below) |
On AMD, Unsloth lists the supported chips by code name: Ryzen AI 300 (Strix Point, Krackan Point), Ryzen AI Max (Strix Halo) and Ryzen AI 400 (Gorgon Point). You also need Windows 11 or 64-bit Linux. The Lemonade Linux guide notes that Ryzen 7000, 8000 and Ryzen AI 200 chips have the older XDNA 1 NPU, which is not supported. For Intel, Microsoft's Foundry Local reference lists NPUs from Arrow Lake (15th Gen) and later with Intel NPU driver 32.0.100.4239 or newer. For Qualcomm, it lists Snapdragon X Elite and X Plus with NPU driver 30.0.140.0 or newer.
Memory matters more than the chip. The NPU shares your system RAM. AMD's walkthrough says the 35B Qwen3.6 model with full context needs about 28 GB available to the NPU, and because Windows by default shares 50% of memory, that means a 64 GB laptop.
Find your starting point
Step by step: Windows 11 with an AMD Ryzen AI chip
- Check that Windows sees the NPU. Press Ctrl+Shift+Esc to open Task Manager, click Performance, and look for NPU in the left column.
- Update the NPU driver. The FastFlowLM project, the AMD runtime Unsloth uses, says to use AMD NPU driver 32.0.203.311 or newer. You can see your version in Task Manager under Performance > NPU, or in Device Manager. Run Windows Update, or get the driver from AMD's driver page.
- Install Unsloth Desktop. Download the Windows app from unsloth.ai/download, or install it from a terminal with winget:
That winget package points at the same Windows installer as Unsloth's GitHub release. The app is labeled beta.winget install --id UnslothAI.Unsloth -e - Turn on the NPU. Open the model picker, click the filter, change it from All to NPU, and click Enable NPU. Unsloth says this one-time setup downloads FastFlowLM and checks your driver, and tells you exactly what to fix if something is missing.
- Download a model. Click one from the NPU list (see the table below for which one fits your RAM). AMD's tip: if the download looks stuck, it is not; let it finish.
- Start chatting. Go back to New Chat, open the model menu, scroll to the NPU section and load your model. Unsloth shows the speed in tokens per second as it answers.
Optional settings, from the gear next to a downloaded model: Power Mode (Power saver, Balanced, Performance or Turbo; FastFlowLM's default is Performance, and Unsloth suggests Power saver or Balanced on battery) and context length (Auto sizes it to your free RAM). Changes apply the next time the model loads.
Step by step: Linux with an AMD Ryzen AI chip
Linux needs three things before Unsloth can use the NPU: the amdxdna driver (built into kernel 7.0 and newer, or added with the amdxdna-dkms package), NPU firmware 1.1.0.0 or newer, and an unlimited memory-lock limit. These commands are from the Lemonade project's Linux guide.
- Ubuntu 24.04 or 25.10: add the Lemonade team's package archive and install the driver, then reboot. (Lemonade's guide lists no driver step for Ubuntu 26.04.)
sudo add-apt-repository ppa:lemonade-team/stable sudo apt update sudo apt install libxrt-npu2 amdxdna-dkms sudo reboot - Check the driver and firmware:
If the first command prints nothing, the driver isn't loaded. If the firmware is older than 1.1.0.0, update yourlsmod | grep amdxdna cat /sys/bus/pci/drivers/amdxdna/*/fw_versionlinux-firmwarepackage. - Raise the memory-lock limit. Open
/etc/security/limits.confwithsudo nano /etc/security/limits.conf, add these two lines, then log out and back in:* soft memlock unlimited * hard memlock unlimited - Install Unsloth Desktop from unsloth.ai/download. The Nov. 19 release includes a Linux AppImage and an Ubuntu
.deb. Then follow steps 4 to 6 from the Windows section: filter to NPU, click Enable NPU, pick a model.
Strix Halo owners: Lemonade warns that the kernel setting amd_iommu=off, which some people use to speed up llama.cpp on the graphics chip, also hides the NPU. Remove it if your NPU doesn't show up.
Prefer the terminal? The Nov. 19 release adds a command that lists NPU models, their size and whether they fit your NPU's memory. Unsloth's guide shows how to serve and chat with a model:
unsloth list-npu-models
unsloth run --model Qwen3.6-35B-A3B-NPU2 --context-length 16384
unsloth chat Qwen3.6-35B-A3B-NPU2
unsloth run prints a local address and API key you can paste into any app that speaks the OpenAI chat format.
Intel Core Ultra and Snapdragon: use Foundry Local
Unsloth's NPU support is AMD-only for now. On an Intel or Snapdragon laptop, Microsoft's Foundry Local is the closest equivalent. It runs models through each chip maker's own software layer: OpenVINO for Intel, QNN for Qualcomm.
- Update the NPU driver. Intel owners: get the Intel NPU Driver for Windows. Snapdragon owners: Microsoft's minimum is Qualcomm NPU driver 30.0.140.0; check Windows Update or your laptop maker's support site.
- Install Foundry Local in a terminal, then check it works:
winget install Microsoft.FoundryLocal foundry --version - See what your NPU can run. The first run downloads the acceleration components (Microsoft calls them execution providers) for your hardware, so give it a minute:
foundry model list --device npu - Chat. Copy a model ID from that list and run it. Using the full ID, not the short alias, makes sure you get the NPU version:
foundry run <model-id-from-the-list>
Microsoft says Foundry Local's catalog changes over time, so we don't list specific models here. If the NPU list is empty, check your driver version against the minimums above.
What about a Mac?
None of these tools run chat models on the Mac's Neural Engine. Unsloth Desktop does run on Macs, but there it uses Apple's MLX and GGUF models, and MLX runs on the CPU and the GPU. Apple Silicon is still very good at local AI, because its memory is shared with the GPU. Our buyer's guide to home AI computers has the Mac picks, and our llama.cpp 1.0 guide covers the engine most Mac setups use.
Which model to pick for your RAM
The NPU runs a special 4-bit format called Q4NX. These are the models Unsloth highlights, with the speed FastFlowLM measured. Unsloth doesn't say which chip produced these numbers, so treat them as a guide, not a promise.
| Model (as named in Unsloth) | Download | Can do | Recommended RAM | Tokens/sec | First token |
|---|---|---|---|---|---|
LFM3-1.7B-NPU2 | 1.5 GB | Reasoning, tools | 16 GB | 40.7 | 0.12 s |
Qwen3-VL-4B-Instruct-NPU2 | 3.9 GB | Vision, tools | 16 GB | 20.5 | 1.05 s |
Qwen3.5-9B-NPU2 | 7.6 GB | Vision, reasoning, tools | 32 GB | 9.5 | 1.11 s |
Gemma4-12B-IT-NPU2 | 8.7 GB | Vision, audio, reasoning, tools | 32 GB | 6.8 | 1.70 s |
Qwen3.8-27B-NPU2 (MTP) | 18.1 GB | Vision, reasoning, tools | 48 GB | 2.0 | 3.80 s |
Qwen3.6-35B-A3B-NPU2 | 20.6 GB | Vision, reasoning, tools | 64 GB | 17.2 | 3.13 s |
"Tokens/sec" is roughly how fast words appear; a token is a word or a piece of a word. "First token" is how long you wait before the answer starts.
| Your RAM | Start with | Why |
|---|---|---|
| 16 GB | LFM3-1.7B-NPU2, or Qwen3-VL-4B-Instruct-NPU2 if you want it to read images | The two models Unsloth recommends for 16 GB. LFM3 is the fastest on the list (40.7 tokens a second). |
| 32 GB | Qwen3.5-9B-NPU2; Gemma4-12B-IT-NPU2 if you need audio | A real step up in answers. Gemma 4 12B is the only model on the list that handles audio, but it is the slower of the two (6.8 vs. 9.5 tokens a second). |
| 48 GB | Qwen3.5-9B-NPU2 for speed; Qwen3.8-27B-NPU2 only for slow, careful work | The 27B model is listed at 2.0 tokens a second, about two words or word pieces a second. |
| 64 GB or more | Qwen3.6-35B-A3B-NPU2 | The largest model on the list, and also one of the quicker ones (17.2 tokens a second), because it is a mixture-of-experts model: the “A3B” in its name means only about 3B of its 35B parameters work on each token. AMD says it needs about 28 GB available to the NPU. |
What you can actually do with it
- Ask about private documents. Unsloth's NPU guide lists chatting with local documents. Contracts, medical letters and tax forms never leave the laptop.
- Read a screenshot or photo. Most of the models above, including the small Qwen3-VL 4B, accept images. Paste an error message screenshot or a picture of a receipt and ask what it says.
- A coding helper that doesn't phone home. Point Cline or VS Code at the address
unsloth runprints. AMD names both among the apps that work. - Chat while the GPU is busy. Keep a model loaded on the NPU while a game, a video export or an image model uses the graphics chip.
- Work on a plane. Set Power Mode to Power saver, load a small model before you board, and you have an assistant with no Wi-Fi.
How to prove it's really local
- Download first, then go offline. Load your model, then turn off Wi-Fi (or unplug the cable). Ask it something. If it answers, it is running on your machine.
- Watch the NPU work. In Task Manager, open Performance > NPU. The FastFlowLM project says this is where you can see NPU usage. The graph should climb while the model writes its answer.
- Turn off web tools. Unsloth chats can use web search and deep research. Those reach the internet by design, so switch them off when you want a fully offline session.
- Know who can connect. Unsloth says its managed model server accepts connections only from your PC, uses a new key each time it starts, and stops when you quit the app.
One honest gap: we haven't audited every network request the app makes, such as update checks. The offline test above is the one that settles whether your chats need the internet.
Honest limits and troubleshooting
- Only XDNA 2 AMD chips. Older Ryzen AI chips are not supported, and neither are Intel or Qualcomm NPUs in Unsloth.
- It's beta software. Unsloth Desktop is labeled beta, and its NPU support is one day old. Expect fast updates; Unsloth shipped two more releases on Nov. 20.
- Not every model runs. The NPU needs Q4NX models, not the GGUF files Ollama and LM Studio use. Unsloth's Nov. 19 release lets you convert a GGUF on its Export page, and says converted models now load on the NPU for text only. (Its guide's FAQ, written earlier, still says this isn't supported yet.) Each NPU model is tagged Runs on NPU, Convert then runs, Converts won't run, or Not supported.
- Big models are slow, and some are very slow. The 27B model on the list is rated at 2.0 tokens a second. The NPU is about efficiency; a good graphics card may well be faster, and we found no official comparison.
- "My NPU isn't showing up." On Windows, update the AMD NPU driver to 32.0.203.311 or newer. On Linux, check
lsmod | grep amdxdna, the firmware version and the memlock limit, and removeamd_iommu=off. Unsloth says setup tells you which of these is missing. - Out of memory? Choose a smaller model or lower the context length in the model's settings. Remember Windows shares only part of your RAM with the NPU by default.
- We didn't test this on our own hardware. Steps and numbers come from Unsloth, AMD, the FastFlowLM and Lemonade projects, and Microsoft. If something on your machine behaves differently, tell us with the contact link at the top of the page.
Keep going
- Buyer's guide: home AI computers, if you're shopping for a laptop or desktop with more memory.
- Open-source AI guide, for the models themselves.
- Mesh LLMs explained, for splitting one big model across several machines at home.
- llama.cpp 1.0, the engine behind most home GPU and CPU setups.
- AI API pricing per 1M tokens, to see what the cloud would charge for the same work.
FAQ
What is an NPU?
A neural processing unit is a chip built into many recent laptop processors to run AI math efficiently. Microsoft's Copilot+ PCs require one rated at 40 or more trillion operations per second (TOPS). AMD says its Ryzen AI 300 series NPU delivers up to 50 TOPS.
How do I know if my laptop has an NPU?
On Windows 11, open Task Manager (Ctrl+Shift+Esc), go to the Performance tab and look for an NPU entry. Microsoft says Task Manager shows NPU usage on devices that have one. Your processor name is the other clue: AMD Ryzen AI, Intel Core Ultra or Snapdragon X.
Which AMD laptops can run Unsloth on the NPU?
Unsloth's guide lists Ryzen AI 300 (Strix Point, Krackan Point), Ryzen AI Max (Strix Halo) and Ryzen AI 400 (Gorgon Point), all with XDNA 2 NPUs, on Windows 11 or x64 Linux. The Lemonade project's Linux guide says Ryzen 7000, 8000 and Ryzen AI 200 series chips have the older XDNA 1 NPU, which is not supported.
Is the NPU faster than my graphics card?
Not necessarily, and we found no head-to-head benchmark from AMD or Unsloth. Unsloth's pitch is efficiency: its release notes say NPUs are 2 to 4 times more efficient per joule per token than GPUs, and running on the NPU leaves your GPU free for games, rendering or video. That is the vendor's claim; we have not measured it.
Can Ollama or llama.cpp use my NPU?
Ollama's hardware support page lists NVIDIA and AMD Radeon graphics cards, Apple Metal and Vulkan. It does not list NPUs. Unsloth's NPU path uses AMD's FastFlowLM runtime and its own Q4NX model format instead of ordinary GGUF files.
Does this work on a Mac?
Not on the Neural Engine. Unsloth Desktop runs on Macs, but its NPU support is for AMD Ryzen AI chips. On Apple Silicon it runs MLX and GGUF models, and Apple's MLX framework runs on the CPU and the GPU. Our buyer's guide covers which Macs are best for local AI.
Do I need the internet?
Only to install the app and download a model. After that, chats run on your laptop. Unsloth says its managed model server accepts connections only from your own PC. Turn off any web search or deep research tool in the chat if you want nothing to leave the machine.
Is it free?
Unsloth calls Unsloth Desktop a free, open-source app; it is still labeled beta. FastFlowLM's code is MIT licensed, and its NPU kernels are closed source but free for personal and commercial use under its own license. Microsoft Foundry Local installs with one winget command.
Sources and method
Show all 18 sources
Published 2026-11-20, updated 2026-11-20. Every figure on this page comes from the linked vendor documentation or release notes, checked Nov. 20, 2026. We did not run these benchmarks ourselves. Sources: Unsloth, AMD NPU: Run Models Locally Guide; Unsloth v0.1.920-beta release notes (Nov. 19, 2026); Unsloth installation guide; Introducing Unsloth Desktop; AMD, Unlocking the World's First NPU for Unsloth Desktop (Nov. 19, 2026); AMD Newsroom, AMD and Unsloth Bring NPU Support to 100+ Agentic AI Applications (Nov. 19, 2026); AMD, Ryzen AI 300 Series processors; FastFlowLM README (AMD ROCm); Lemonade, LLMs on Linux with FastFlowLM; Microsoft Learn, Foundry Local CLI reference; Microsoft Learn, Get started with Foundry Local; Microsoft, Windows 11 and Copilot+ PC requirements; Microsoft Learn, Copilot+ PCs developer guide; Microsoft Support, Copilot+ PC experiences; Intel NPU Driver for Windows; Ollama, Hardware support; Apple MLX README; Phoronix, Unsloth Desktop Ryzen AI NPU support on Linux (Nov. 19, 2026).