OGAM - releases

Everything that ships to your phone, in the order it shipped. Two tracks:

Stable - the tested build on the App Store and Google Play. What you get if you just install the app.
Beta - new features first, weekly, from GitHub releases. Expect the occasional rough edge.

Current App Store release: iOS 0.0.103. For Android, see the current Google Play release. New beta builds and the full per-build history live on GitHub releases.

0.0.103 - iOS stable - July 17, 2026

Now live on the App Store. Real control over memory, and a hard pass over the things that were breaking.

Choose how hard your phone works. New loading modes - Lean, Balanced, Aggressive - let you decide how much memory models are allowed to use.
See and free your RAM. The model manager shows what’s loaded and how much each model is holding, with a per-model eject and an Eject All that frees everything, voice and transcription included.
Bigger models, no crashes. 9B models load and run fast on iOS and Android, and switching models is memory-aware, so a new one loads without running out of RAM. Over budget? Load Anyway measures your real free memory first.
Voice you can trust. Voice notes and dictation transcribe reliably and send as text the model can read, spoken replies no longer read tool-call text aloud, and stopping speech stops cleanly.
Downloads that recover. Failed and app-killed downloads come back as retryable cards instead of vanishing, and queued downloads survive a restart.
Stopping keeps your reply. Stop mid-stream and you keep what was written, reasoning included - no more mislabeling a stopped turn as a token-limit cutoff, and remote LM Studio and Ollama reasoning now shows in the Thinking block.

Your phone’s real hardware, put to work.

GPU and NPU acceleration, on by default when your device supports it, so models run faster without you touching a setting.
Smart model residency decides what fits in memory - and a “Load Anyway” override lets you push past the limit when you know better, for the rest of the session.
Clearer guidance when a model quietly falls back to CPU, with a one-tap path to a faster build.

Downloads that behave, memory you can see.

A download queue that won’t freeze the app - concurrent downloads are capped, queued items show a clear “Queued” state, and everything survives an app kill.
Per-process memory in Device Info - Available, Footprint, and Process Limit, so you can see exactly what your phone has to spare.
Voice-note transcription rendered right in the sent message.
A top-level error boundary so a render error recovers instead of leaving you on a white screen.

Tools, membership, and routing that thinks.

MCP tools, routed on-device through an embedding model, so the right tool is picked without a cloud round trip.
Off Grid AI Pro membership with license-key activation and a device cap.
Smart routing that tells text from image requests and warms the right models at boot.
Voice mode that speaks image-generation results back to you.

Looking for an older build or the exact commit list? Every version, with full notes and downloads, is on GitHub releases.