Loading…
AGI Mobile · iPhone & Android · Coming soon
Your phone runs the model and keeps the chat.
On an iPhone with Apple Intelligence the app generates through Apple's Foundation Models runtime, and on Android it generates through AICore. Otherwise you download AGI Standard once and it runs through ExecuTorch or llama.rn. Either way the thread is written to SQLite on the phone.
- iOS · Apple Foundation Models
- Android · AICore
Coming soon
Download on theApp Store
Get it onGoogle Play
On-device inference
A Local answer is generated by the phone you are holding.
The app picks a runtime from what the hardware actually reports (the system model, a downloaded one, or nothing yet) and it tells you which before you send anything.
The model the operating system already holds
On an iPhone where Apple Intelligence is available, AGI generates through Apple's Foundation Models runtime. On an Android device with AICore, it generates through Google's on-device model. Neither one is downloaded, and the app checks whether the runtime is available, downloadable, downloading, or unavailable before it offers you the model.
A model you download once
Choose AGI Standard instead and the phone loads a 4B Qwen3 build, roughly 2 GB quantized, with a 262,144-token context. It runs through ExecuTorch where the device reports at least 3.5 GB of RAM, and through llama.rn from the downloaded file where it does not.
A picker that says which one is loaded
Every row in the model picker carries one of three states (ready, download required, or locked) so the cost of a tap is visible before you make it. Generation also pauses on its own while the device reports thermal throttling.
What tapping Cloud does
Switch to AGI Cloud?
Sign in to use AGI Cloud chat. Your local chat stays on this device unless you choose to start a Cloud session.
That is the dialog, word for word. Local and Cloud sit side by side in the chat screen, onboarding lands you on Local with no account, and Cloud stays behind that prompt until you sign in. Sending chats you already have is a separate action under Settings → Data Controls: it copies conversation titles and message text, leaves file attachments and memory facts on the phone, and runs once when you press it.
In the app
The drawer carries the whole workspace.
Local chatThreads run and stay on the device, and the model picker sits in the composer beside the attach and voice controls.Projects and recentsOne drawer holds chats, projects, artifacts, the library, connectors, skills, schedules, and tasks.Memory and personalizationMemory facts are held on the phone and are excluded even when you choose to sync chats to AGI Cloud.Data controlsExport chats, memory, settings, and installed model details from the device, and wipe local-only chats from Storage.
Build sheet
Here is what the build contains.
| Default model | AGI Standard · Qwen3 4B Instruct · ~2 GB quantized · 262,144-token context |
| Lighter option | AGI Lite · 1B on-device build · ~1.1 GB |
| Chats and memory | SQLite on the device |
| Export | Runs on the device: chats, memory, settings, installed model details |
| Cloud sync | Manual and one-time · attachments and memory facts stay on the phone |
| App lock | The Face ID, Touch ID, or passcode already enrolled on the phone |
| Model training | Off. Prompts, responses, and files are not used to train AGI-owned models |
| Platforms | iPhone and Android · Coming soon |
Public launch: date to be announced
AGI Mobile is not in the app stores yet.
The runtimes, the drawer, and the settings above are built and running on device. Neither store listing is published, and no release date has been announced.
iPhone & Android · To be announced