Version History

Changelog

Everything that has shipped, and what is coming next.

v1.4
April 2026
Latest Release
Current Release

Pro Features arrive. A record-breaking 24B model, Audition Mode for side-by-side model comparison, AI Photo & Video Enhancer, and more — all on-device, all private.

Pro Features
  • Audition Mode — send one prompt to every installed model and compare results side by side.
  • AI Photo & Video Enhancer — upscale to 4K on-device using the Neural Engine.
  • Cinematic Grade — three film-grade colour profiles for upscaled video.
  • Advanced GPU Tuning — unlock models above 10B parameters.
  • Mistral Small 3.2 24B, running natively on iPhone 17 Pro, entirely on-device.
  • Advanced performance metrics and diagnostics.
Also In This Update
  • Performance improvements across inference and UI.
  • Bug fixes and stability improvements.
v1.3
April 2026
Public Release

New model, eight languages, and meaningful accessibility improvements. All free, all on-device.

Models
  • Nemotron 3 Nano added. NVIDIA's hybrid Mamba-2 and Transformer architecture gives it a 262,000 token context window with far lower memory use than comparable models.
Language Support
  • The app is now available in English, French, Spanish, Dutch, German, Simplified Chinese, Traditional Chinese (Hong Kong), and Korean.
  • The AI automatically responds in your device language. No configuration required.
Accessibility
  • VoiceOver support across all screens with descriptive labels and hints throughout.
  • Reduce Motion respected — animations suppressed when enabled in iOS Settings.
  • Dynamic Type supported at all accessibility size categories.
  • Response completion announced to VoiceOver users when generation finishes.
Stability
  • Context window budget now based on effective device capacity rather than theoretical model limits.
  • Prompt decoding batched correctly — no more stalls on long conversation histories.
  • EXAONE Deep reasoning output filtered cleanly before rendering.
v1.2
30 March 2026
Previous Release

The moment LiberaGPT went public. Free to download, no account, no subscription, no data leaving your device.

Fixes and Improvements
  • Memory management across long conversations made more robust.
  • Switching between models is cleaner with no leftover state.
  • Several stability issues under heavy use resolved.
  • Minor performance improvements throughout.
v1.1
March 2026
App Store Submission

The build submitted to Apple for review. Some housekeeping was needed before it could go live.

Models
  • Phi-4 Mini added. Strong reasoning and instruction following in a compact package.
  • EXAONE Deep added. Particularly good at maths, logic, and working through complex problems.
  • AceInstruct added. Fast and precise, built for quick focused answers.
  • Thinking mode added for models that support it. The app shows when a model is working through a problem before responding.
Performance
  • The app now monitors device temperature in real time and adapts continuously. You get strong responses without the phone running hot.
  • Live performance stats available for those who want them: generation speed, thermal state, and memory at a glance.
  • Visible warnings when the device is under thermal or memory pressure, with a clear option to start fresh.
Output
  • Responses now render with proper formatting: headings, lists, code blocks, and more.
  • Output stays readable while streaming, even mid-sentence.
Other
  • Diagnostics screen added under Settings. Device health, model info, storage, and error logs all in one place.
  • Fast and Deep generation modes available. Fast prioritises speed, Deep prioritises quality.
  • Auto-routing added as an option. The app selects the appropriate mode based on your question.
  • Time, date, and approximate location can be shared with the model as optional context.
  • Conversation history view added. Tap the clock icon to browse, resume, or delete past conversations.
  • Copy, undo, and stop controls added to the chat interface.
  • Full appearance settings added: light and dark mode, font size, animations, and haptics.
  • Knowledge Packs removed. It was not ready.
v1.0b
Early 2026
Closed Beta

A small group of testers got their hands on it via TestFlight. Their feedback shaped what shipped.

What testers experienced
  • A working AI assistant, running fully offline on their iPhone. No server, no subscription, no catch.
  • SmolLM3 ready out of the box. StableLM Zephyr available to download.
  • Conversations saved locally and persistent across sessions.
  • Streaming responses with the ability to stop generation at any time.
  • A full onboarding flow and mandatory terms acceptance before first use.
v1.0a
Early 2026
Internal

The first build that worked. Used internally at 5N6 LTD to prove the concept was real. It was. We kept going.