In March 2026 Microsoft quietly announced it was rolling Copilot out of several Windows 11 apps — Photos, Notepad, Snipping Tool and the Widgets panel are all losing their AI bolt-ons. The official line is that the company wants to refocus on "meaningful" AI integration. The unofficial reading is that nobody was using these features and a lot of people were complaining about them.
Either way, it's a good moment to ask: what do you actually want AI doing on your PC? And do you want it phoning home to a cloud you don't control every time you ask it a question?
I'm Rai — I build free Windows software from Singapore. Most of what I build is offline by design. So this list is opinionated. Every tool here runs locally, and most of them don't even need an internet connection after install.
What's actually being removed
The TechCrunch report from March 2026 confirms Copilot is being pulled from:
- Photos — the AI image-edit panel that nobody asked for.
- Notepad — the "rewrite this" sidebar most users disabled.
- Snipping Tool — the "ask Copilot about this screenshot" prompt.
- Widgets panel — the AI summary cards.
Microsoft is doubling down in two specific places: File Explorer (Copilot in the right-click menu) and Edge (full Copilot UI redesign rolling out June 2026). So they're not killing Copilot — they're concentrating it. Which means if you don't want Copilot in Edge, this is a good time to switch browsers anyway.
Why local AI is having a moment
Two things changed in 2025–2026 that made local AI viable for normal users instead of just researchers:
- Open-weights models caught up. Llama 3.3, Qwen3, Mistral Small 3, and a dozen others now run well on a 12 GB GPU — and even on CPU if you're patient. The quality gap to GPT-class models is narrow for most everyday tasks.
- The tooling got friendly. You used to need Python, CUDA setup, and a willingness to debug DLL errors. Now you install one app, click "Download model", and you're chatting locally.
The trade-off is honest: cloud AI is faster, has no hardware requirements, and currently still wins on the absolute frontier of model quality. Local AI keeps your data on your machine, costs nothing per query, and works offline. For a lot of people that's a better deal.
1. LM Studio — easiest way into local LLMs
If you've never run a model locally before, start here. LM Studio is a free Windows app that gives you a model browser (browse Hugging Face from inside the app, click download), a chat interface that looks like ChatGPT, and a local OpenAI-compatible API server you can point other apps at. Works on CPU and any modern GPU.
Realistic spec needed: 16 GB RAM and a recent NVIDIA card with 8 GB+ VRAM gets you running 7-8B parameter models at usable speed. 12 GB+ VRAM unlocks the 12-14B class models that feel close to GPT-4-mini quality.
2. Ollama — same thing, more developer-friendly
If you're comfortable with a terminal, Ollama is what you want. ollama run llama3.3 downloads the model and drops you into a chat. It exposes a local API on port 11434 that any app can talk to, and there's a growing ecosystem of front-ends that connect to it (Open WebUI is the popular one for browser-based chat).
Same hardware needs as LM Studio. The choice between them is mostly aesthetic — LM Studio if you like clicking, Ollama if you like typing.
3. RBS Voice Cloner V2 — local voice AI
Disclosure: I make this one. RBS Voice Cloner V2 is built on the open-source XTTS v2 engine. Bundles PyTorch and CUDA 12.8 inside the installer so you don't have to set anything up. After the first model download (~2 GB), everything runs locally. 16 built-in voices, unlimited custom clones from a 30-second sample, 17 languages, 7-band parametric EQ.
Use case: voiceovers, audiobook drafts, language practice, accessibility (turning long articles into audio you can listen to on the train). If you're paying ElevenLabs and feeling the pinch, this fills most of the gap.
4. Stable Diffusion + ComfyUI — local image generation
Stable Diffusion has been around for years now and the open-weights ecosystem is far bigger than Midjourney's curated catalogue. ComfyUI is the front-end most people land on in 2026 — it's a node-based interface that looks intimidating for five minutes and then makes total sense. Free. Runs on any 8 GB+ VRAM card. SDXL Turbo gives you Midjourney-class results in 2-3 seconds per image.
5. Whisper.cpp — local transcription
OpenAI's Whisper model, repackaged in C++ for speed and small enough to run anywhere. Drag a meeting recording onto it, get a full transcript in a fraction of real-time. The "large-v3" model is more accurate than most paid transcription services. Faster-Whisper is a similar fork that runs on CTranslate2 and is even faster on GPUs.
6. AnythingLLM — local chat with your documents
What Notion AI does, but on your machine. Drop PDFs, Word docs, and webpages into a workspace, let it index them, then ask questions. The model and the index both stay local — no document ever leaves your PC. Pairs naturally with LM Studio or Ollama for the underlying chat model.
7. ImageGlass + free local upscalers (Real-ESRGAN)
If you used Photos' Copilot mainly to upscale or clean up images, the open-source replacements are noticeably better. Real-ESRGAN runs locally and produces 4x upscales that hold up on a phone screen. Several free GUIs wrap it (Upscayl is the popular one). Pair with ImageGlass as your photo viewer and you've replaced the bits of Photos that Copilot was bolted onto.
How to actually disable Copilot in the apps it's still in
Microsoft is removing Copilot from those four apps over the course of 2026, but the timeline is staggered and there's no "off" switch in the UI. If you want it gone today:
- Notepad / Photos / Snipping Tool: uninstall the latest update via Settings > Apps > Installed apps. The pre-Copilot versions still work fine.
- Widgets: right-click the taskbar > Taskbar settings > turn off Widgets entirely. Most people who don't use Widgets won't miss it.
- Edge: Settings > Sidebar > turn off Copilot. Or switch to Firefox / Brave / Vivaldi if you don't trust the redesign coming in June.
A practical local AI stack for 2026
If you're starting from zero and want to set up a sensible privacy-respecting AI workflow on Windows, here's what I'd actually pick:
- Chat / writing help: LM Studio with Llama 3.3 70B (if you have the VRAM) or Mistral Small 3 (if you don't).
- Voice / TTS: RBS Voice Cloner V2 for cloning, or Piper if you just want fast TTS in a known voice.
- Image generation: ComfyUI with SDXL Turbo.
- Transcription: Whisper.cpp.
- Document Q&A: AnythingLLM pointed at your LM Studio instance.
Total cost: zero. Total telemetry: zero. Privacy: as good as your OS allows.
Worth saying out loud
Local AI is not always better than cloud AI. If you genuinely use Copilot in Edge for research and it's saving you time, keep using it. The point of this post isn't "Microsoft bad, local good". It's that the AI features being removed weren't doing much, and the people who actually want AI on their PC have better, freer, more private options now than they did a year ago.
If you try any of these and run into trouble, send me an email through the contact page. I read every message myself.