The NPU used to be the throwaway spec on the sticker. That is finally changing. Snapdragon X Elite laptops shipped in 2024 with a 45 TOPS NPU, Intel followed with Lunar Lake, and AMD with Ryzen AI 300, so a Copilot+ PC now has a piece of silicon designed to run models locally, quickly, and without cooking the battery. Softonic’s coverage of the next CPU generation from Arm, Intel, and AMD confirms the trend, on-device inference is a real target now, not marketing.
The catch is that not many apps use the NPU yet. Windows Studio Effects and Recall use it. A handful of third-party apps have caught up. This list covers the eight desktop apps worth installing on a Copilot+ PC to actually use the NPU you already paid for.
What to look for in a Copilot+ PC app
Not every “AI” app on the Microsoft Store uses the NPU. A lot of them ship the same GPU or CPU-based pipeline as before and rebranded the marketing. What matters:
- Native NPU execution. The app explicitly targets QNN (Qualcomm), OpenVINO (Intel), or DirectML NPU. Otherwise inference falls back to CPU and drains battery.
- Arm64 native builds. On Snapdragon X hardware, x64 emulation still works but with a real hit to performance and battery. Arm64-native releases are non-negotiable.
- Local-only mode. The whole point of NPU acceleration is to keep sensitive input on-device, no round-trip to a cloud API.
- Model size discipline. NPUs on current Copilot+ PCs peak around 45 TOPS. That is plenty for 7B-parameter language models and small vision models, not for 70B behemoths.
Quick comparison
| App | Best for | Platforms | Free plan | Starting price | NPU acceleration |
|---|---|---|---|---|---|
| LM Studio | Chatting with local LLMs | Windows Arm64, x64, Mac | Full app free | Free | Native on Snapdragon X |
| Ollama | Running LLMs from CLI or API | Windows Arm64, x64, Mac, Linux | Free | Free | DirectML on select hardware |
| Krisp | Real-time voice cleanup | Windows, Mac | 60 min/day free | About $8/mo | NPU voice model on Copilot+ PC |
| Adobe Lightroom | AI photo enhancement | Windows, Mac | 7-day trial | Around $10/mo Photography plan | Enhance uses NPU on Copilot+ PC |
| DaVinci Resolve | AI video editing | Windows, Mac, Linux | Full free tier | $295 one-time Studio | Neural Engine on NPU-capable hardware |
| Camo Studio | AI webcam effects | Windows, Mac | Free tier | Around $40/year Pro | Background segmentation on NPU |
| CapCut | AI-driven video edits | Windows, Mac | Full free tier | About $8/mo Pro | On-device caption and remove |
| Whisper.cpp | Local speech-to-text | Windows, Mac, Linux | Free, open-source | Free | Qualcomm QNN and OpenVINO backends |
The apps
1. LM Studio -- Best for chatting with local LLMs
LM Studio is the fastest way to get a local model chatting on a Copilot+ PC. The Arm64 build ships an NPU runtime that uses Qualcomm’s QNN backend, so 7B and 8B parameter models like Llama 3, Phi, and Mistral run at usable speed with the fan off. The app itself is a chat window plus a model catalogue, no config needed.
Where it falls short: the NPU backend only accepts models that have been converted for QNN. The catalogue is growing but still smaller than the general GGUF library. If you have a specific fine-tune in mind, check that it exists in NPU-ready form before you install.
Pricing:
- Free: the full app, all features
- Paid: none
Platforms: Windows Arm64, Windows x64, macOS
Download: LM Studio for Windows/Mac
Bottom line: the single most impressive use of an NPU on a Copilot+ PC. Install this first.
2. Ollama -- Best for local LLMs from the command line
Ollama is the tool developers and scripters use to run models from a terminal or an API. It exposes a REST endpoint on localhost, which every serious IDE integration and coding assistant knows how to hit. On Copilot+ PCs, DirectML support is landing progressively, so recent builds tap the NPU on supported hardware instead of falling back to CPU.
Where it falls short: the CLI-first design is not for everyone, and the NPU path is younger than LM Studio’s. Expect some models to run on GPU or CPU today rather than NPU, depending on the runtime version.
Pricing:
- Free: the app and every core feature
- Paid: none, open-source
Platforms: Windows Arm64 and x64, macOS, Linux
Download: Ollama for Windows/Mac/Linux
Bottom line: the right pick if you want to script against a local model or plug one into an IDE. LM Studio has the nicer UI, Ollama has the API.
3. Krisp -- Best for real-time voice cleanup
Krisp is one of the earliest third-party apps that switched its noise-suppression model onto the NPU on Copilot+ PCs. It sits in the audio path and cleans microphone input in real time, so Zoom, Teams, OBS, and every DAW get a filtered signal without any per-app configuration. On battery, moving the model off CPU has a real, measurable effect.
Where it falls short: the free tier caps at 60 minutes of noise cancellation per day. And Krisp is not an editor, only a filter.
Pricing:
- Free: 60 minutes per day
- Paid: about $8 per month unlimited
Platforms: Windows, Mac
Download: Krisp for Windows/Mac
Bottom line: turn it on, forget it, and your calls sound cleaner. On a Copilot+ PC, the battery cost is close to zero.
4. Adobe Lightroom -- Best for AI photo enhancement
Adobe Lightroom ships a set of AI-driven features, Enhance (super-resolution), Denoise, and masking, that used to hammer the GPU. On a Copilot+ PC, recent Lightroom updates route these through the NPU where possible, so a batch of RAW files finishes faster and the fan stays quiet.
Where it falls short: it is still Adobe. Subscription-only, Creative Cloud login required, and photographers who resent Adobe billing already know this. Also, only some AI operations use the NPU today; other steps continue to run on GPU.
Pricing:
- Free: 7-day trial
- Paid: around $10 per month on the Photography plan bundled with Photoshop
Platforms: Windows, Mac
Download: Adobe Lightroom for Windows/Mac
Bottom line: for photographers already inside the Adobe subscription, the NPU acceleration is a free speed-up. For anyone else, this is not the right entry point.
5. DaVinci Resolve -- Best for AI video editing
DaVinci Resolve has a Neural Engine that powers speech-to-text, magic mask (object segmentation), voice isolation, and scene detection. On Copilot+ PCs, recent Studio builds run some of these on the NPU, and Blackmagic has been steadily converting more of the model set. The base app is free, and the NPU work happens automatically.
Where it falls short: the biggest Neural Engine features, better AI upscaling and specific masking modes, are Studio-only. And Resolve is famously heavy, install size and RAM appetite are large.
Pricing:
- Free: full base app
- Paid: $295 one-time for Studio
Platforms: Windows, Mac, Linux
Download: DaVinci Resolve for Windows/Mac/Linux
Bottom line: for anyone editing video on a Copilot+ PC, the free version already benefits. Studio unlocks the rest.
6. Camo Studio -- Best for AI webcam effects
Camo Studio turns a phone or a webcam into a proper virtual camera with framing, background segmentation, and lighting effects. On Copilot+ PC hardware, the segmentation model shifts to the NPU, so a Zoom call can run a portrait-mode background without the CPU spiking or the fan spinning up.
Where it falls short: the Pro tier gates the good effects, and Camo works best paired with an iPhone or iPad as the input source. Casual users on integrated webcams get less benefit.
Pricing:
- Free: basic capture
- Paid: around $40 per year for Pro
Platforms: Windows, Mac
Download: Camo Studio for Windows/Mac
Bottom line: the fix if your webcam looks terrible on video calls. On battery, NPU segmentation makes it a lot less costly.
7. CapCut -- Best for on-device video editing
CapCut has quietly become the video editor most creators actually use, and its Windows desktop build shifts more of its AI features (auto-captions, background removal, style transfer) onto the NPU on Copilot+ PCs. That means a 20-minute video can be captioned or masked without opening a cloud tab.
Where it falls short: ByteDance owns CapCut, which raises data-handling concerns for some editors. The free tier has watermark and export caps that push users to Pro.
Pricing:
- Free: full editor with export limits
- Paid: about $8 per month CapCut Pro
Platforms: Windows, Mac
Download: CapCut for Windows/Mac
Bottom line: the pick for short-form and social video editors who want fast AI features on-device.
8. Whisper.cpp -- Best for local transcription
Whisper.cpp is a lightweight C/C++ port of OpenAI’s Whisper transcription model. On Copilot+ PCs it can target Qualcomm’s QNN or Intel’s OpenVINO backend, which pushes the model onto the NPU. The result, a full-quality transcription pipeline that runs entirely offline, faster than realtime, on a laptop.
Where it falls short: no polished UI. Whisper.cpp is a set of binaries and Python bindings. Downloading models and running commands is the entry cost.
Pricing:
- Free: open-source, all features
- Paid: none
Platforms: Windows, Mac, Linux
Download: Whisper.cpp on GitHub
Bottom line: the transcription pipeline for anyone who cares about privacy or offline capability. Wrap it in a script and it becomes a batch transcriber.
How to pick the right one
If you just bought a Copilot+ PC and want to see the NPU do something impressive right now, install LM Studio. Full stop.
If you code and want a local model that IDE plugins can hit, Ollama is the better fit than LM Studio because of its REST endpoint.
If your NPU laptop is a work machine and video calls dominate the day, Krisp and Camo Studio together clean up audio and video with almost no battery cost.
If you edit photos, Adobe Lightroom’s Enhance and Denoise are the largest NPU wins in Adobe’s Creative Cloud, worth the subscription if you already shoot RAW.
If you edit video, DaVinci Resolve free tier is enough for a lot of NPU benefit, and Studio unlocks the rest.
If you make social-format short video, CapCut is a faster on-ramp than Resolve.
If you record interviews or long-form audio and want transcripts, Whisper.cpp produces the best offline result. Willing to script a little.
FAQ
What is a Copilot+ PC and how do I know if I have one? A Copilot+ PC is a Windows 11 laptop with an NPU rated at 40 TOPS or higher. Snapdragon X Series, Intel Core Ultra 200V (Lunar Lake), and AMD Ryzen AI 300 all qualify. Check Settings, About, for “Copilot+ PC” or look at the sticker.
Do these apps require Copilot+ hardware? No, all eight run on other Windows PCs and most on Mac too. The NPU acceleration only kicks in on hardware that has one. On a regular laptop, the same apps fall back to GPU or CPU.
Which apps use the NPU without any configuration? LM Studio, Krisp, DaVinci Resolve, and CapCut detect the NPU automatically. Ollama and Whisper.cpp need the right backend selected. Lightroom and Camo Studio route selectively depending on the operation.
Are Snapdragon X (Arm64) apps as fast as Intel or AMD x86 versions? Arm64-native builds are competitive and often faster on the NPU. x64 apps running under emulation are slower and use more battery. The apps in this list all ship Arm64-native versions.
Can I run models bigger than 7B on the NPU? On current 45-TOPS NPUs, 7B and 8B parameter models are the sweet spot. 13B works but at reduced speed. 70B is not realistic on today’s NPU; a GPU with a lot of VRAM is still the answer for those.
Do NPU apps work offline? Yes, that is the appeal. LM Studio, Ollama, Whisper.cpp, DaVinci Resolve’s Neural Engine, and Krisp all run their models locally without any network traffic.
