Local AI on your Mac: questions and answers

Local AI means running an AI model on your own device instead of on a company's servers. You download the model's files once. Then your computer does the work, so your prompts and files are answered right there, even with no internet connection. Nativ, the local AI app for Mac, does this on Macs with Apple Silicon.

Nativ is a free, open-source local AI app for Mac. It runs open AI models on Apple Silicon, and nothing leaves your Mac. You can chat with AI models, dictate, take meeting notes, create images and connect your coding tools, all with models running on your own Mac. Nativ is made by Neywa Labs.

A local LLM is a large language model (the kind of AI that reads and writes text) that runs on your own computer. Labs such as Google DeepMind, Qwen and Meta publish open models that anyone can download, often on Hugging Face. An app like Nativ loads one of them on your Mac, so you can chat with it with no server in between.

The difference is where the model runs. With cloud AI, your prompt travels to a company's servers and the answer comes back over the internet. With local AI, the model runs on your own computer, so the prompt and the answer stay there, and it works offline. The trade-off: a local model has to fit in your computer's memory, so the models you can run depend on your Mac.

Yes, by design: a local model reads your prompts and files on your own computer, not on a server. An app can still go online for other things, such as model downloads, updates or usage data, so check what it sends. Nativ's privacy policy says Neywa Labs does not collect your chats, prompts, files or answers (see what leaves your Mac).

Only to download the model. Once the model is on your computer, it answers with no connection at all. In Nativ, you need the internet to download models and app updates. Once the models you use are on your Mac, chat, dictation, transcription and image generation work offline.

Yes, if your Mac has Apple Silicon (M1 or later) and runs macOS 26 or later. Download Nativ, pick the model it suggests for your Mac and ask your first question. There's no server to set up and no account to create.

Nativ runs models with MLX, Apple's open-source machine learning framework, which is built for Apple Silicon chips (M1 and later). On these Macs, the processor and the graphics chip share one pool of memory (unified memory), so a model can use most of your Mac's memory. MLX doesn't run on Intel Macs, so Nativ doesn't either.

Three steps:

1. Check your Mac: open the Apple menu, then About This Mac. You need an Apple chip (M1 or later) and macOS 26 or later.
2. Download Nativ from runnativ.ai, drag it into Applications and open it.
3. Download the model Nativ recommends for your Mac, then ask your first question.

Then turn off your Wi-Fi and ask again. It still answers.

A Mac with Apple Silicon (M1 or later) on macOS 26 or later, enough memory for the model you choose, and free disk space for the model files. You don't need an account, a server or a subscription. A free Hugging Face token is needed only for gated models, whose makers ask you to accept their terms before you download.

Start with the model Nativ recommends when you first open it: it's chosen for your Mac's hardware. Once that's working, open Models to try others. Nativ warns you before you download a model that's unlikely to fit in your Mac's memory.

Nativ installs like most Mac apps:

1. Go to runnativ.ai and click Download for macOS.
2. Open the downloaded file (for example Nativ-0.4.4.dmg) and drag Nativ into your Applications folder.
3. Open Nativ. If macOS asks whether to open an app downloaded from the internet, click Open.
4. Follow the first-launch steps: pick the recommended model or one you already have, optionally set a server API key, and allow microphone, accessibility and screen recording access now or later in Settings.
5. Ask your first question.

Prefer Terminal? Homebrew installs the same app (see the next question).

Yes. Run brew install --cask nativ in Terminal. Homebrew installs the same app from the same release, and it needs Apple Silicon and macOS 26 (Tahoe) or later. After that, Nativ updates itself from inside the app.

Every Nativ release is signed by the team and notarized by Apple: Apple scans it for malicious software before macOS lets it open. The full source code is public on GitHub under the MIT license, so anyone can read what the app does. Download Nativ only from runnativ.ai, its GitHub page (github.com/Neywa-Labs/nativ) or Homebrew.

No. You get Nativ from runnativ.ai, from its GitHub releases or with Homebrew. Nativ is made by Neywa Labs, and its source code is at github.com/Neywa-Labs/nativ.

Nativ updates itself. It checks for new versions and installs them from inside the app, so you don't need to download it again. Release notes for every version are on GitHub at github.com/Neywa-Labs/nativ/releases.

Quit Nativ, then drag it from Applications to the Bin. Models are stored separately, in your Hugging Face cache or the folder you chose in Models, so delete them in Models first if you want the disk space back. If you installed with Homebrew, brew uninstall --zap --cask nativ also removes Nativ's settings and app data.

A model needs about as much memory as its file size, plus some extra for long conversations. In our tests of Recursive Flint, for example, the 23 GB 4-bit model peaked at about 23 GB of memory, and a 128,000-token prompt (a very long document) added about 4 GB. Pick a model well under your Mac's total memory, so macOS and your other apps have room. Nativ warns you before you download a model that's unlikely to fit.

Each model is a separate download, from a few gigabytes to tens of gigabytes. The models on our home page, for example, are 3.2 GB, 10.3 GB and 19.4 GB, and the Nativ app download itself is under half a gigabyte. Nativ checks for free space before each download, and you can keep models on an external drive formatted as APFS (Models, then Sources, then Choose External Location).

More than 5,000 open models, for chat, coding, vision, speech, image generation and more. They come from Hugging Face in MLX format, which is built for Apple Silicon, and you can browse and download them inside Nativ, in Models.

Often on the day they're released. Nativ has shipped launch-day support for new open models from labs including Google DeepMind, Cohere, Meta, Qwen, Z.ai, OpenBMB and Gradium. The Nativ blog has posts with measurements on Apple Silicon for several of them.

No. Nativ is a Mac app for Apple Silicon (M1 or later) on macOS 26 or later. It doesn't run on Intel Macs, Windows and Linux.

Yes. Once the models you use are on your Mac, Nativ runs them with no internet connection: chat, dictation, transcription and image generation all keep working. You only need the internet to download models and updates. So when the Wi-Fi gives up an hour into a flight, your chat just keeps going.

Your chats, prompts, files and answers stay on your Mac, and Neywa Labs does not collect them. Nativ goes online for two things. It downloads the models you pick, straight from Hugging Face, where they're hosted. And it checks for app updates. The full details are in the privacy policy.

No. Neywa Labs does not use your chats, files or other content to train, fine-tune or evaluate AI models, or for advertising, and it does not sell personal information. Your content is processed on your Mac, so Neywa Labs doesn't receive it in normal use.

No. Nativ has no sign-up, no account and no subscription. The only optional step is a free Hugging Face token, needed just for gated models (models whose makers ask you to accept their terms first).

Each permission turns on a feature, and you can skip any of them and turn it on later in Settings. Microphone: to hear you when you dictate or record a meeting, asked the first time you record. Accessibility: so the dictation shortcut works in other apps and the text lands where your cursor is. Screen recording: to capture system audio when you record a meeting, and to attach a screenshot to a chat. Both stay on your Mac. Dictation recordings stay on your Mac, and their raw audio is deleted after five minutes.

Nativ puts chat, voice and images in one app on your Mac. You can chat with AI models about text, images, video and your documents, dictate into any app, take meeting notes with a transcript and summary, and have answers read aloud. You can also create and edit images, and connect your coding tools to a model running on your Mac. It all runs locally.

Yes. Record a meeting and, when you stop, Nativ writes a transcript and a summary on your Mac, with no account and no calendar to connect. You can also import audio files to transcribe and summarize them. To keep a copy, export the summary and transcript together as PDF, DOCX or TXT from Audio, then Library.

Yes. Tap Fn and Control together to start, speak, then tap them again to stop: the text appears at your cursor in whatever app you're using. A speech model on your Mac does the transcribing, so it works offline. End with the word "enter" and Nativ presses Return for you.

Yes. Open the Images tab, pick an image model and describe what you want. Nativ creates and edits images with models running on your Mac, so your prompts and pictures stay on it.

Yes. Nativ serves the model running on your Mac through OpenAI- and Anthropic-compatible APIs (the standard ways apps talk to AI models), so coding tools can use it like any other provider. The Integrations page sets up Claude Code, Codex, Pi, Hermes, OpenCode, Aider, Goose, Crush, Qwen Code and OpenClaw for you. Editors such as VS Code, Cursor, Zed, JetBrains IDEs, Cline and Continue can connect too.

Yes. Nativ runs a local server at http://127.0.0.1:8080 with OpenAI-compatible endpoints (/v1/chat/completions, /v1/responses, /v1/models, plus image and audio routes) and the Anthropic-compatible /v1/messages. No key is set by default; setups that need one use nativ. You can change the host, port and key on the Developer page.

Yes. Nativ finds compatible MLX models in your Hugging Face cache (it follows the HF_HUB_CACHE and HF_HOME settings) and in any extra folders you add. Models you already have show up without a second download.

MLX is Apple's open-source machine learning framework for Apple Silicon. mlx-vlm is an open-source package for running and fine-tuning vision language models on Mac with MLX, and mlx-audio is an open-source library for text to speech, speech to text and speech to speech built on MLX. Neywa Labs builds both, and Nativ bundles an mlx-vlm server, so the app and the library share one engine.

Yes. Nativ is free, with no account and no subscription, and it's open source under the MIT license. The full source code of the Mac app is on GitHub at github.com/Neywa-Labs/nativ, so you can read it, build it yourself or contribute.

Neywa Labs makes Nativ. Neywa Labs is a local AI company. It builds the open-source engine and apps that run open AI models natively on Apple devices: mlx-vlm, mlx-audio and Nativ. The company is based in Kraków, Poland.

Ask in the Nativ Discord, open an issue on GitHub (github.com/Neywa-Labs/nativ/issues), or reach the team through runnativ.ai/contact.