Private AI That Runs Entirely
On Your Device.
Whisper runs powerful language models fully on your device. No servers, no accounts, no tracking. Talk to it hands free, show it images, and let it act on your phone: calendar, alarms, contacts, messages. Offline by default. Everything stays yours.
Runs the open models you already love
- Qwen
- Gemma
- Llama
- Phi
- SmolLM3
It doesn't just answer. It acts.
Ask for something and a small model running on your own phone picks one of 18 tools wired into Android, fills in the arguments, and runs it.
Calendar & time
4 toolsReads your schedule, and books into it when you say so.
People & messages
4 toolsFinds the contact, drafts the message, hands you the send button.
Web
4 toolsThe one place it reaches outside your phone, and only on request.
Your device
6 toolsThe small things you would otherwise dig through settings for.
Anything that creates something on your phone stops and asks for a tap first. Everything else is a read, and reads never leave the device.
▸All 18 tools by name
- list_calendar_events
- create_calendar_event
- schedule_reminder
- set_alarm
- search_contacts
- dial_number
- compose_sms
- compose_email
- web_search
- web_fetch
- open_url
- open_maps
- get_location
- get_battery
- read_clipboard
- write_clipboard
- set_brightness
- search_phone_media
Ask for something…
Talk to it. Hands free.
Open Live and just speak. Whisper listens, works out when you have finished, answers out loud, and reopens the mic. The whole loop runs in airplane mode, with nothing streaming anywhere.
01It hears you stop
Voice-activity detection watches the mic level and ends your turn on a second of silence. No wake word, no push-to-talk.
02It transcribes, on device
whisper.cpp turns the audio into text right on the phone. The recording never becomes a file that leaves it.
03It thinks, on device
The same local model you picked answers, and can reach for any of the 18 tools while it does.
04It speaks back
Kokoro-82M synthesizes the reply through sherpa-onnx, then the loop reopens the mic for your next sentence.
Eight voices to answer in
Preview each one in Settings and keep the one you like. Every voice is synthesized on the phone.
- Ariadefault
- Bella
- Nicole
- Michael
- Fenrir
- Puck
- Emma
- George
You can watch it think.
Small models are the ones most likely to claim they did something they didn't. Whisper is built so that claim is structurally impossible, and so you can check it yourself.
It cannot fake success
Every planning turn is forced through a grammar that only accepts a real decision: call this tool with these arguments, or answer. A model that would otherwise reply “I've set your alarm!” without setting one literally cannot emit that.
Decisions aren't rolled for
Planning is decoded greedily at temperature zero, because it is control flow, not prose. Only the sentence you actually read is sampled, so the choice to act doesn't change from one run to the next.
You can read the log
Turn on the trace and every decision, tool call, argument and result is recorded step by step, raw model output included, with an export button. Nothing about the loop is hidden from you.
Nine models. Or bring your own.
Download, switch and delete models inside the app. Whisper reads your phone's actual RAM and free storage first, and warns you before you pull something that would get killed by the OS.
Qwen3 4B Instruct
Strongest at phone actions. The one to pick for the assistant.
Gemma 4 E2B
Multimodal. The one that can look at your screenshots and photos.
Llama 3.2 3B Instruct
Meta's compact chat model. Fast, friendly, reliable for everyday questions.
Phi-4 Mini Instruct
Microsoft's small model. Strong at math, logic and structured answers.
SmolLM3 3B
Fully open, reasons step by step. A good balance of speed and smarts.
Qwen3 1.7B
Smallest and fastest, and it shows its thinking. For older phones.
Qwen3 1.7B · Uncensored
Refusal removed at the weights, not prompted around. For older phones.
Qwen3 4B · Heretic
Optimised abliteration: refusals gone, the model otherwise unchanged.
Gemma 4 E4B
The bigger sibling. Needs a 12 GB device, and the app will tell you.
Any GGUF on Hugging Face works
Paste a direct model URL on the Models screen and it downloads like any built-in one, progress bar and all. Add the matching mmproj file and that model gets vision too. The catalog is a starting point, not a walled garden.
No servers, because there is no backend.
Whisper has no accounts, no analytics, no telemetry and nothing to sign in to. There is no server to send your data to even if someone wanted to. Here is the exact line, in both directions.
Passes the airplane test
Switch the radios off and everything still works: the chat, the voice loop, the calendar, the alarms. That is the whole proof, and you can run it yourself.
No server on the other end
There is no Whisper backend. Nothing to breach, nothing to subpoena, nothing that goes down. Your phone talks to itself.
Nothing to sign in to
No account, no email, no phone number. Install it, pick a model, start talking. There is no profile of you anywhere because none is ever created.
No analytics, no telemetry
No usage tracking, no crash reporter, no advertising ID. There is no analytics SDK in the app, so there is nothing to opt out of.
- Analytics SDKnot included
- Crash reportingnot included
- Advertising IDnot included
- Session replaynot included
Your data stays where it is made
Every message, recording, transcript, contact lookup and calendar read lives on the phone, along with the models and the full agent trace.
- Messages
- Voice
- Contacts
- Calendar
- Location
- Photos
- Models
- Trace
Leaves only when you ask
Exactly three things ever go out, each behind your own tap, and nothing else rides along with them.
- you tapA web search queryonly the query, when you ask
- you tapA page URL to fetchwhen you ask it to read one
- you tapA model downloadonce, from Hugging Face
Before you download
Does it really work with no internet?
Yes. Once a model is downloaded, chat, live voice, calendar, alarms, contacts and the rest run in airplane mode. The only two tools that need a connection are web search and page fetch, and they only run when your request needs them.
Why isn't it on the Play Store?
It is not listed yet. For now it is a free APK you install directly, and Android will ask you to allow installs from your browser the first time. Nothing about the app changes when it does reach the store.
Which model should I start with?
Qwen3 4B Instruct if you want the assistant that actually does things on your phone, since it is the strongest of the small models at picking tools. Gemma 4 E2B if you mostly want to show it images. On a 4 GB phone, start with Qwen3 1.7B.
How fast is it?
Honestly: a 2-4B model on a phone CPU is usable, not instant. Expect a short wait for a reply, longer on older hardware. Smaller models are noticeably quicker, which is why the library spans 1 GB to 6 GB rather than shipping one size.
Will it fill up my phone?
Models are between 1 GB and 6 GB and none of them ship inside the app, so you choose what to download and can delete any of them from the Models screen. Whisper shows your free storage and RAM before you start a download, and warns you if a model would not fit.
Can I run a model that isn't in the list?
Paste any direct GGUF URL from Hugging Face on the Models screen and it downloads like a built-in one. Add the matching mmproj file alongside it and that model gets vision as well.
Is there an iPhone version?
Not yet. Android is first; the codebase is Expo and already builds for iOS, so it follows once the assistant is proven on Android.
Your assistant, on your phone, and nowhere else.
Free, no account, no sign-in. Install it, pick a model, and put the phone in airplane mode to see for yourself.