Off Grid - Private AI Chatv0.0.103
Released July 16, 2026Major
The latest major update drove a 1.06 star improvement in the Android rating average, moving from 2.27 to 3.33. Despite this gain, the release introduced critical stability regressions and user concerns regarding model output, resulting in a volatile reception. The net impact is neutral as functional gains are currently offset by technical friction.
What changed
- PerformanceIntroduced Model Loading modes (Lean, Balanced, Aggressive) for granular memory control.
- New featureAdded a Model Manager to monitor RAM usage and eject specific models.
- New featureIntegrated Transcription (STT) and Text-to-Speech (TTS) into Model Settings.
- UXAdded a toggle to enable or disable 'Thinking' for reasoning models directly in chat.
- UXAdded cloud markers to distinguish between remote and on-device models.
- PerformanceImproved download resilience with retryable cards and persistence across app restarts.
- UXEnhanced voice note reliability and added a slide-to-cancel UI for push-to-talk.
- PerformanceOptimized 9B model loading speed and implemented memory-aware model switching.
- UXAdded image generation guides for GPU trade-offs and fixed resolution issues.
- SocialAdded social links for X and Slack and a 'Stay in the loop' card in Settings.
Reception
New feature
The addition of model management tools and local integration received positive feedback for enabling easier plug and play functionality.
Performance
backfiredPersistent initialization hangs on device analysis screens prevent app access, contradicting the performance improvement intent.
How it played out
The Android rating average improved by 1.06 stars following the latest update, reflecting a shift toward polarized user reception. While the introduction of model management tools and improved local integration garnered praise, new technical regressions including initialization hangs and perceived model censorship have introduced significant friction. Momentum appears volatile as users weigh functional improvements against new stability and content-filtering concerns.
Marlvel takeaways
- Address initialization hangs on device analysis screens to resolve the primary barrier to app access.
- Investigate user reports of model censorship and reduced text coherence to restore trust in generation quality.
- Correct UI text scaling issues on high-density displays to improve readability for affected Android users.
Original release notes
This is a big stability and control release. New: - Model Loading modes (Lean, Balanced, Aggressive) to control how much memory models get - Model manager shows what's loaded and its RAM, with per-model Eject and Eject All (voice and transcription included) - Transcription (STT) and Text-to-Speech (TTS) now in Model Settings - Turn Thinking on or off for reasoning models right from chat - Cloud marker distinguishes remote models from on-device ones - Repair action for vision models missing their projector file - "Stay in the loop" card in Settings, plus Follow on X and Join Slack links Fixes & improvements: - Downloads interrupted by an app-kill come back as retryable cards instead of vanishing; queued downloads survive a restart and resume - Failed downloads show in the badge count and Download Manager - Voice notes and dictation transcribe reliably; push-to-talk has a clear slide-to-cancel pill and the mic stays usable while a model loads - Stopping a reply keeps what was already written, reasoning included, and no longer wedges the chat or mislabels a stopped turn; a genuine token-limit cutoff now shows a clear indicator - Thinking toggle applies to the current turn; remote reasoning (LM Studio, Ollama) now shows in the Thinking block - Broken or partly-downloaded models are caught before generation with a plain error instead of a silent failure - 9B models load and run fast; switching models goes through the memory-aware loader so it doesn't run out of RAM - Load Anyway is offered for any model over budget and measures real free memory after clearing space - Image generation guides you on the GPU speed/quality trade-off and no longer produces garbage-resolution output - Image settings use your values, and Reset to Defaults resets them too - Context-length slider goes up to each model's real trained maximum - Knowledge-base search only runs when you turn it on - The image viewer closes before the Save prompt, and audio releases before the photo picker, fixing input and voice-mode hangs - Faster first launch, and Aggressive mode no longer overcommits memory