KernelAI

Open models.On your phone.

53 open models from 15 labs, running entirely on your phone. No account, no subscription, and your chats stay on your device.

A KernelAI conversation in which Gemma 4 E2B identifies a photo of the Great Pyramid of Chichén Itzá, offline

Open weights from fifteen labs

  • Qwen
  • Google
  • Community
  • Hugging Face
  • Meta
  • Stability AI
  • IBM
  • Microsoft
  • DeepSeek
  • Liquid AI
  • PrismML
  • Mistral
  • 01.AI
  • NVIDIA
  • OpenAI
Runs on your phone.
Not on someone else's computer.

WHAT YOU CAN DO

A whole assistant, with the network off.

Documents

Read a document that does not fit.

Attach a PDF, a Word file or a spreadsheet and ask about it. Nothing is uploaded.

KernelAI answering questions about an attached CSV of quarterly financials

Vision

Point it at a photo.

Photograph a sign, a menu, a chart or a receipt and ask what it says.

KernelAI identifying a dog's breed from a photo

Models

Choose from 53 local models.

Browse by provider and download what fits: 112 MB to 3.11 GB, from Qwen and Google to PrismML and DeepSeek.

The KernelAI model catalog grouped by provider, from Meta and Google to Mistral AI

Code

Code anywhere, even offline.

Ask for a function, a fix or an explanation and get syntax-highlighted code back, with no connection.

Qwen 2.5 Coder 3B writing a syntax-highlighted Python binary search

Web search

Look things up, when you choose to.

Off by default. Your own Tavily or Exa key turns it on, and every answer comes back with its sources.

KernelAI answering a web search with inline cited sources

Your own model

Bring your own model.

Import a GGUF from Files or a direct URL and run it alongside the catalog. Experimental, and entirely yours.

The import screen testing an imported GGUF model for compatibility

Custom instructions

Set the rules once.

A system prompt shapes every conversation: tone, format, persona. Presets get you started.

The system prompt editor with its presets

Voice

Speak naturally. Stay private.

Dictation runs on the device too, so the question you ask out loud goes nowhere else.

On-device dictation listening in a conversation

Shortcuts

Run local AI from Shortcuts.

Ask KernelAI and Extract Text from Document are Shortcuts actions, so any automation can use the models on the phone.

A Summarize Documents shortcut chaining Extract Text from Document into Ask KernelAI

THE PART NOBODY ELSE DOES

We tell you what your phone can run before you download it.

Every model card is measured against your device's real memory budget. Models that will not fit are disabled, with the numbers printed on them:

Requires 3.4GB RAM Your device: 2.9GB

The catalog runs from 112 MB to 3.11 GB, and 18 of the 53 models are under a gigabyte.

See the full catalog

Now on Android

Private AI, now on Android.

Download KernelAI from Google Play, choose a model that fits your device, and keep your chats on it.

Get it on Google Play

WHAT PEOPLE SAY

Reviews from the App Store.

There are more on the listing, including the unkind ones.

  • AdequatePipe
    “Running on an iPhone 14, it’s smooth, fast, has multiple chats, multiple models, works offline”
  • sayjak
    “On your phone for safety free to use multiple different kind of bots to do different kind of things”
  • Nackylacky
    “Cant lie a great ai it does everything i need to do”

QUESTIONS

Common questions

  • How does on-device AI work?

    KernelAI runs quantised open-weight models through llama.cpp, on your iPhone's CPU and GPU. After a one-time download over Wi-Fi the model file lives on your device, and every prompt is processed locally with no internet round-trip.

  • Is it really free?

    Yes. Current features are free, with no subscription, no usage caps and no account. Optional tips support development and do not unlock features. Models run on your hardware.

  • What leaves my device?

    Your chats, attached files, images and transcriptions are processed on the phone. KernelAI does not upload the full conversation to a server of ours. Model downloads and optional web search use the network; the privacy policy explains every supporting service.

    Read the privacy policy
  • Does it work offline?

    Completely. Once a model is downloaded you can use it in airplane mode, underground, or anywhere with no signal.

  • Which devices are supported?

    iPhone and iPad running iOS 16.4 or later, and Android devices running Android 7.0 or later. KernelAI measures your device's memory and disables the models it cannot run, so you always see what will actually work.

  • Can I talk to it?

    Yes. Dictation uses the iPhone's own on-device speech recognition, so the audio is transcribed on the phone and never sent anywhere. The app can also read replies aloud with the system's voices.

  • Does it work with Siri and Shortcuts?

    Yes. Two Shortcuts actions, Ask KernelAI and Extract Text from Document, run entirely on the phone, so an automation can answer a prompt or pull the text out of a document without opening the app. Siri can trigger either one.

  • Can I use my own model?

    Yes. Import any GGUF file from Files or straight from a URL, including a server on your own network, and it appears beside the catalog models. Imports are marked experimental, because compatibility varies from file to file.

  • Can it search the web?

    Yes, with your own key, and it is off by default. Add a Tavily or Exa API key and the models tagged Web Search can look things up and cite their sources. The key stays in the iOS Keychain and the query goes to the provider you chose, not to us.

  • How does it compare to ChatGPT?

    KernelAI runs open models locally on your phone and needs no account. Ordinary chat works after downloading a model, without sending the full conversation to our servers. Optional web search sends a short query to your selected provider; the privacy policy explains that network access.

Put it onyour phone.

GET THE APP

Free, no account, and it works with the network off.

Download on the App StoreGet it on Google Play