The same M-series chip that edits your video also runs language models, draws images, and transcribes speech, all on a single shared pool of memory. Most people rent three separate subscriptions to do those things on someone else's servers while that hardware sits idle. Off Grid AI Desktop is a free, open-source app that runs chat, image generation, and voice directly on your Mac.
Free, open source, runs offline. No account.
This is the broad tour: chat, images, and voice. Image generation is the heaviest part, so size your Mac for that.
| Tier | Chip | RAM | macOS | Free disk |
|---|---|---|---|---|
| Minimum | M1 / M2 | 8 GB | 13 Ventura | 20 GB |
| Recommended | M2 Pro / M3 | 16 GB | 14 Sonoma+ | 40 GB |
| Comfortable | M3 Max / M4 | 32 GB+ | 14+ | 60 GB+ |
Chat and voice run fine on 8 GB. Image models want more headroom, so 16 GB is the comfortable floor if you generate pictures often.
One app covers the three things you would otherwise pay three services for. Everything below runs on your own silicon.
A ChatGPT-style window backed by a model on your drive. Temperature and context controls are right there. The model's HTML, SVG, Mermaid, and React output renders live in a sandboxed preview.
Built on stable-diffusion.cpp
. Run SDXL, the fast few-step SDXL-Lightning, SD 1.5 and 2.1, or the 2026 flagship Z-Image-Turbo at around 8 steps. You get text-to-image and image-to-image, a live per-step preview, progress and ETA, a cancel button, a lightbox, an artifacts gallery, and style presets like Sketch, Cinematic, and Anime.
Voice in uses bundled whisper.cpp
: tap the mic and your speech becomes text in the composer. Voice out uses Kokoro-82M, an open-weight multilingual model, for a per-message Speak button and an auto-speak mode that reads replies aloud.
None of this calls home. You can draft a prompt, paint an image, and have the answer read to you on a flight with the Wi-Fi off.
Three engines run side by side on Apple Silicon. llama.cpp
handles chat, stable-diffusion.cpp
handles images, and whisper.cpp
handles speech. Kokoro produces the spoken voice.
Each one runs through Metal, Apple's GPU framework. Because the chip uses unified memory, the GPU reads model weights from the same RAM the CPU uses, with no copy across a PCIe bus. That is why a fanless laptop can chat, draw, and listen without a separate graphics card. The models are quantized so they fit in consumer RAM and stay fast.
A few habits keep all three engines responsive.
Cloud AI services send your words, your images, and your voice to a remote server, where they are logged and may train the next model. With Off Grid AI Desktop, all of it stays on your Mac. The prompt, the picture, the recording: none of it leaves the machine.
No account. No telemetry. No API key. The app is AGPL-3.0, so the full source is on GitHub for you to read. There is nothing to log in to and nothing phoning home.
git clone https://github.com/off-grid-ai/desktop
cd desktop
npm install
npm run dev
Yes. Free and open source under AGPL-3.0. Chat, image generation, and voice are all in the free open core.
Yes. It uses stable-diffusion.cpp
through Metal on your GPU. The image never touches a server.
Yes. After the models download, chat, images, and voice all run with no network.
SDXL, SDXL-Lightning, SD 1.5 and 2.1, and Z-Image-Turbo, with text-to-image and image-to-image.
8 GB handles chat and voice. 16 GB is comfortable once you add image generation.
Yes. Speech is transcribed on-device by whisper. The audio and the text stay on your Mac.
Chat, draw, and talk to AI on your own Mac today, with nothing leaving your drive.