The easiest way to run local AI on your own computer.
Install in one click. Chat, code and make images with open models that run on your computer's own GPU, on a Mac or a Windows PC. Every dial is within reach, and nothing sits between the model and your hardware to slow it down.
Also on Mac (Apple silicon, macOS 14+)
Also on Windows (10/11 64-bit, beta)
or install from Terminal
curl -fsSL https://getsabine.com/install | bashor install from PowerShell
irm https://getsabine.com/install.ps1 | iexSwitch between and in the sidebar, just like the real app. Code mode shows Sabine running tools on a project.
The idea
Easy by default. Powerful when you reach for it.
Most local-AI tools make you pick one: a friendly app with the knobs hidden, or a pile of flags and servers you assemble yourself. Sabine is built to refuse that trade.
Installs itself
Open the disk image (Mac) or run the installer (Windows). On Windows the installer contains everything, Python included, so nothing else needs to be installed. Pick a model, or let Sabine pick one that fits your computer.
Nothing in the way
The model runs in-process, straight on your GPU: Apple's MLX on a Mac, llama.cpp on Windows. No local server to proxy through, no wrapper re-reading your whole prompt every turn.
Yours to turn
Model, voice, content policy, thinking budget, permissions, spend limits, memory. The simple path is the default; every control is one click deeper.
Speed
Fast because it does less work, not because it cuts corners.
On a local model the slow part is usually not generating, it's re-reading. Sabine is engineered around not doing that.
Where one reply's time goes
A 200-token reply on an M-series, 16 GB Mac, from Sabine's own measurements.
Re-reading the conversation every turn 32 s
Sabine: attention cache kept between turns 16 s
Bars are drawn to scale of the figures in Sabine's README. Your numbers depend on your chip, model and prompt; the stats line under every reply shows the real ones.
Measured, not promised
Sabine ships with an eval harness that runs real tasks through the app's real chat loop and grades the result.
- Native engines. MLX on Apple silicon (4-bit weights, unified memory); llama.cpp with GGUF models on Windows.
- Thinking has a budget. A hard cap on reasoning tokens, and none at all for a quick “thanks”.
- Tuned per mode. Code gets a lean prompt and low temperature; chat keeps its personality.
Control
Two dials. Every voice at every level.
Personality is how Sabine talks. Policy is what it will answer. They're independent, so you can have a blunt answer that still flags real stakes. Drag them and see.
Levels 1–4 share one hard floor. The settings are real; the replies on the right are an illustration of tone.
Off until you say so
File access, web search, a real browser, image generation: every capability is a switch that starts off, per mode. While it's off, the model isn't even told it exists.
Approve or deny
When the model wants to do something sensitive, it stages a request. Approve or deny it right in the chat, on the record.
A model per mode
Chat and Code each keep their own model, voice and switches. Use a small fast model for chat and a bigger one for code.
Incognito chats
Chats that keep nothing: not the text, not the pictures. Local files from them are cleared when the server starts.
Features
Everything in one real desktop app.
A native window (with a Dock icon on Mac), the same interface in a browser, and a terminal chat if that's how you live.
A model that fits your computer
Tell Sabine your hardware and it recommends the best local model for chat and for code, and warns you when one fits but would crawl. 43 models to browse.
Local by default, online when you choose
One conversation, two engines. Flip to a frontier model through OpenRouter with your own key, with the same prompt, personality and commands.
Weights on your disk. Nothing leaves the machine to run the model.
A Code workspace
Reads, writes and edits files, runs your tests and fixes what fails, in a project folder you choose. Asks before it touches anything.
Skills
Expert playbooks the model can load, in the same SKILL.md format Claude Code uses. A UI/UX one ships built in; add or import your own.
Image studio
Prompt, aspect, how many, and a searchable gallery. Run FLUX or Z-Image locally (MLX on Mac, stable-diffusion.cpp on Windows), or any image-capable model on OpenRouter.
Memory that's one paragraph you can read
Sabine learns lasting facts as you talk and brings them into every chat. It's a single paragraph: paste in a ChatGPT or Claude memory, edit it, or wipe it. Duplicates are dropped automatically.
⌘K for everything
Every action, setting, model and conversation in one fuzzy search.
Spend tracking and a hard credit limit
Using online models? Every message is priced, counted in a ledger, and stopped at a daily, weekly or monthly cap you set. Local is always free.
The internet is opt-in
/online adds search and page fetching. /browser gives it a real Chromium that keeps your logins. Both start off, and you can narrow them to the sites you allow.
Signed updates
Updates download in the background and install on quit. Each is cryptographically signed; Sabine refuses anything that doesn't verify.
Attach, edit, export
Drag in files, edit a message to rewind the conversation, pin chats, export as Markdown. Syntax-highlighted code, offline.
App, browser or terminal
~/Sabine/run for the CLI, run web for the browser. One engine, three doors.
Models
Open-weights models, ready to run.
Pre-quantized to 4-bit (MLX weights on Mac, GGUF on Windows), so a 3-billion-parameter model fits in about 1.7 GB instead of 6. Search them, filter by family, show only what fits.
Privacy
What you type stays on your computer.
The weights sit on your disk and your GPU does the work. No API key, no account, and the server binds to your machine only.
- Works with the Wi-Fi off. The interface has no build step and loads nothing from a CDN, no web fonts, no remote icons.
- The network is a choice. Only installing and downloading models use it, plus whatever online features you switch on.
- Accessible by design. Text meets WCAG AA contrast, and every looping animation respects reduced motion.
Install
From download to chatting in a few minutes.
No Homebrew, no virtual environments, no config files.
On Mac
Drag Sabine to Applications
Open Sabine.dmg and drag the Sabine icon onto the Applications folder beside it.
Open Sabine
The first launch sets itself up, fetching Python and the libraries (about 600 MB, a few minutes), with a progress window.
Pick a model
Sabine opens. Choose one in Settings; its weights download the first time you do.
shasum -a 256 Sabine.dmg.On Windows (beta)
Run SabineSetup.exe
Open SabineSetup.exe. It installs for your user only, so no administrator password is needed.
Let it set itself up
The installer contains Python and everything Sabine runs on, so it works on a PC with nothing installed. At the end it downloads only the AI engine (llama.cpp, roughly 20–650 MB for your graphics card), and Sabine can fetch it later with one button if you are offline. Models download separately.
Pick a model
Sabine opens. Choose one in Settings; its weights download the first time you do.
Get-FileHash SabineSetup.exe in PowerShell with the SHA-256 in SabineSetup.exe.sha256.Prefer a command line?
Paste the line for your system. On Mac it downloads the app, checks its SHA-256, puts it in Applications and opens it, and macOS doesn't ask about installs done this way. On Windows it runs the same installer from PowerShell.
Mac · Terminal
curl -fsSL https://getsabine.com/install | bashWindows · PowerShell
irm https://getsabine.com/install.ps1 | iexRequirements
Mac
Windows (beta)
FAQ
Questions
Is it as smart as ChatGPT or Claude?
No, and it doesn't pretend to be. Local models are a fast, private helper for drafting, summarizing, explaining code and quick questions. They can lose the plot on long multi-step reasoning. When you need more, switch the same conversation to a frontier model with one key.
Does it run on Windows?
Yes, as a beta on Windows 10 (1809+) and 11, 64-bit. Local models run on llama.cpp (GGUF) instead of MLX, with CUDA on NVIDIA GPUs, Vulkan on others, or the CPU. Chat, Code mode, online models, sandboxed documents, local images and computer use (beta) work the same. The installer isn't code-signed yet, so SmartScreen asks you to click More info, then Run anyway.
Where does it install?
On Mac, Sabine's code, Python environment, settings and chats live in ~/Sabine, and the app is Sabine.app in Applications; to remove it, delete both. On Windows it installs for your user only, with no administrator rights.
Does my data leave my computer?
Not in local mode. Only the installer and model downloads use the network. If you turn on online mode, search or browsing, those specific requests go out.
How do I update?
Sabine checks for updates itself, downloads them in the background and installs them the next time you quit. Each update is cryptographically signed, and Sabine refuses anything that doesn't verify. You can also check by hand in Settings → Updates. Your settings, memory and chats are never touched.
Prefer the terminal?
On Mac, after installing, ~/Sabine/run starts the command-line chat, and ~/Sabine/run web serves the same interface in a browser.
Is the content policy a lock?
No. It's a prompt, and Sabine says so plainly. Levels 1–4 share a hard floor (no routes to chemical or biological agents, mass-casualty weapons, working malware, or sexual content involving minors). “Off” sends no prompt at all, so it has no floor. What each level is worth is measured in the repository's policy-eval.
Put your GPU to work.
Free to download. Install in minutes. Nothing leaves your machine.