• New app languages: Arabic, Greek, Romanian, Portuguese, and Russian • Pick assistant reply language separately from the interface language • Improved on-device tool calling and assistant chat experience • Faster, more stable MLX and GGUF inference • Refreshed settings and UI polish

New: Conversation history drawer Your saved chats now live in a slide-out side panel. Open it with the = button (top-left) or a swipe in from the left edge. Tap a chat to reopen it; swipe a row to delete or export. Tap the dimmed area to close. New: On-device error analysis Settings → Privacy → "Analyze errors with AI" feeds the app's local diagnostic report to your loaded model and explains likely causes of any crashes/errors — fully on-device, nothing uploaded. Polish Cleaner header, larger tap targets, and better text contrast on the light and dark themes. Please test *Open/close the history drawer with both the = button and the left-edge swipe; confirm rows are tappable and the panel is bright (not dimmed) * Run "Analyze errors with AI" after a model is loaded * General stability when switching models

Fixes *Qwen3-VL no longer errors when you attach a photo ("placeholder tokens does not match number of frames" is gone). * Reasoning models (Qwen3, R1, QwQ) now finish their answer instead of getting cut off mid-thinking. If one ever does run out of room, you'll see a clear note instead of a blank reply. New: Tool calling (optional) *Turn on Settings → "Allow tools". The model can now call on-device tools to answer: calculator, memory search, date/time, unit/currency conversion, and OCR on an attached image. Web search is also available (this is the only tool that uses the internet). * Works on tool-aware models (Qwen3, Llama 3.1) and GGUF models. What to try 1. Attach a photo and ask Qwen3-VL to describe it — should work after a multi-turn chat. 2. Ask Qwen3-4B a hard question — confirm an answer follows the reasoning. 3. Enable "Allow tools", then ask: "what time is it", "convert 5 miles to km", or "search the web for something". Note: with tools on, replies take a little longer (the model checks tools first). Everything stays on-device except web search.

Thanks for using TokenAI! Here is what we changed in this build: Check out our sister app! We've added a spotlight card for "On Device : Local AI Studio"inside the settings page. Discover offline voice, chat, and vision options in seconds. *Native App Store Ratings: We've updated our feedback prompt to use Apple's native StoreKit dialog, keeping rating requests non-disruptive and secure. *Bug Fixes: Refined the settings interface to work flawlessly with our latest localized themes.

This update makes every model you load run better — especially ones outside the built-in catalog: • Fixed garbled characters in emoji, Turkish, Chinese and other non-English replies — characters split across tokens now always render correctly. • Long chats now stop cleanly at the model's context limit instead of failing mid-reply with a generation error. • More vision models (LLaVA, MiniCPM-V, Gemma-style and others) now use their native chat template for image and video questions — noticeably better answers. • Downloads are verified the instant they finish: gated-repo error pages and corrupt model files are rejected with clear guidance instead of being saved as broken models. • Memory safety: speculative decoding now checks free RAM before loading a draft model, and switching models mid-generation frees the old one immediately. • Regenerating a response now keeps attached video clips.

This update fixes image and video generation crashes: • Fixed the app fully quitting during image generation — most often with SDXL-class models, and especially while Low Power Mode was on. The memory budget now correctly accounts for Low Power Mode (which lowers the app's memory ceiling), so heavy models no longer start a job they can't finish. • Fixed a crash at the very end of generation (the final image decode) on high-RAM devices: the large model is now always freed before the memory-heavy decode step, instead of staying loaded on top of it. • Image and video generation are more memory-stable overall — GPU memory is reliably handed back between steps and between video frames, keeping peak usage flat across a whole clip. • Added a "Reset image model limits" button in Settings to re-enable any image model that was automatically disabled after an out-of-memory crash.

Pick models per type from Settings (tap any model row) Voice mode button for hands-free conversation Experimental: Gemma 4 12B (text, 16 GB iPad Pro) Fixed: Qwen3-VL "Couldn't Load Model" Fixed: Assistant list no longer shows vision-only models Fixed: composer model chips fit on their own row

What's New in 1.1.0 • Smarter on-device memory: download dedicated embedding models (MiniLM, BGE, nomic) for noticeably better recall and document search — still 100% offline. • Tools the assistant can use: enable "Allow tools" to let it run an on-device calculator and search your saved notes to ground answers. • More reliable, consistently-formatted responses. • Privacy: optional sensitive-data redaction strips emails, card numbers, and other PII before anything is saved to local memory. Everything runs on your device. No account, no cloud.

On-device creativity, upgraded: • New image models: Stable Diffusion 1.5 and DreamShaper 8 — lightweight options that run on 4 GB iPhones — plus DreamShaper XL Turbo for high-end devices. • Text-to-Video (beta): turn a prompt into a short animated clip, generated entirely on your device. • Edit a photo: attach an image in Image mode and describe the change. • Generation controls: quality, negative prompt, up to 4 images per prompt, and video frames, speed & motion. • Reproducible results: lock a seed to recreate a favorite. • Live progress with a cancel button for image and video generation. Fixed: an issue that could stop some image models from loading. Plus stability & performance improvements.

What's new in 1.0.1 • Smoother Dynamic Island camera — pulling down now reveals the live viewfinder instantly, with no black gap during the open animation. • Faster overall — quicker launch, smoother scrolling and menu transitions. • Improved Turkish support — the assistant now reliably understands and replies in Turkish. • Easier to read — higher-contrast chat bubbles and Dynamic Type support on more screens. • Clearer recovery — a Retry option now appears if a model fails to load. • Stability and polish fixes throughout.