This update makes model downloads dependable: start one, close the app, and it is still there when you come back. IMPROVEMENTS - Downloads Survive Closing the App Model downloads now pick up where they left off after the app is closed. When you reopen it, the screen that started the download shows its progress and Cancel button again, and the model is set up as soon as the download finishes. - Resume After Swiping the App Away A model download stopped by swiping the app away in the app switcher now continues from where it stopped with one tap on Resume. The import button in My Models reads Resume Download with the model's name and opens the paused download. BUG FIXES - Download Live Activity Stays Visible The download progress on the Lock Screen and in the Dynamic Island now stays up after the app is reopened, instead of disappearing while the download keeps running. - Hugging Face Imports Finish Reliably A model downloaded from Hugging Face is now added to My Models even if the app was closed during the download. - Retrying Audio and Image Downloads Retrying an MLX audio or image model download no longer deletes a download that had already finished. - Partial Whisper Downloads A Whisper model that only partly downloaded no longer shows as Downloaded. It now offers to continue the download instead.

FEATURES - Neural Engine models: a fourth on-device engine and the only one that keeps generating in the background, including in the AI Keyboard. Five models. (iOS 27) - Apple Private Cloud Compute: Apple's larger cloud model, free within your Apple allowance, no account and no API key. (iOS 27) - Apple models can see your pictures, up to four per message. (iOS 27) - Models without vision get an on-device description of a photo that has no writing in it. (iOS 27) - Core AI models from HuggingFace, filtered to the builds your device can run. - Speech recognition on the Neural Engine: 25 European languages, keeps transcribing in the background. (iOS 27) - 28 Kokoro voices in 756 MB, half the disk of the existing pack. (iOS 27) - Extract Text: one tool for slide decks, spreadsheets, PDFs, e-books, charts and screenshots, from a file or a web address. - Swedish, fully translated, with keyboard layout, prediction, dictation and voices. - Live speech recognition in 33 languages, written down as you speak. - Sub-agents on remote models, each with its own context, model and report. - Python: scipy and scikit-learn for statistics and machine learning. - Python: statsmodels for regression and forecasting, with seaborn. - Python: TA-Lib, backtesting.py and quantstats for market analysis. IMPROVEMENTS - Transcribe Media prefers a video's published captions, and says which source it used. - Text inside a page's images can be read, and those images load in parallel. - Scripts get an app tool's full result, saved to your workspace when too large. - New Home screen: twelve tiles for every chat surface and mini app. - Sequential Thinking retired; Critical Thinking stands alone and frees context. - iCloud Drive chats: lists open at once, cloud-only files are marked not hidden, and a failed read never overwrites good data. - llama.cpp 0.4.1: 310 changes, per-chip attention, Metal 4 on M5 and A19, leak fixes. - MLX 20260918: reliable tool-call parsing, faster ParoQuant loading, own thread. - mlx-audio 20260918: thirteen new speech model families, faster speaker separation. - FluidAudio 20260918: Kokoro fully on the Neural Engine, clearer audio, fuller transcripts. - Scripts can open a file by name from the chat workspace, not just save one. BUG FIXES - Second-device projects and chats no longer look deleted before iCloud fetches them. - File search reports the right line number when blank lines precede the match. - File names with spaces open normally. - Retired tools are removed on update; your selection moves to their replacement. - Long data series no longer come back empty, and the reply says how many it shows. - A stalled reply respects your time limit. - A batch of edits is no longer lost to one malformed tag. - A failed request no longer counts as a completed Agent turn. - A large result saved into the workspace is now complete. Open Apple Models to try Private Cloud Compute, or run a Neural Engine model and switch apps while it writes.

BUG FIXES - Running Python Works Again An urgent fix. Running Python did not work at all in 3.1.0. The agent and skills rely on it to do real work, so this restores a large part of what they can do. Update now and your agents and skills can run Python again.

FEATURES - Doc Scanner Turn a PDF, a photo, or a camera scan into clean text you can read, search, share, or chat with, on device. Each page is analysed twice: a layout model finds every element and its reading order, then each is read on its own with a prompt matched to what it is, so a table stays a table and an equation an equation. Reading a whole page at once is what makes an AI invent text, so Privacy AI does not. An interrupted scan resumes where it stopped. Pick Standard or Thorough detail, whether to keep the page's pictures, and whether to keep page numbers, all changeable afterwards without re-scanning. IMPROVEMENTS - Doc Scanner reads more accurately: the model takes its best guess, not a random one of its ten best, and faint text is no longer discarded unread. Agreement with a real paper's own words rose from 93 to 97 percent - A title repeated on several pages is rendered as a title on each of them, so a multi page form keeps its heading. Figure, chart, and table captions are unchanged - The on-device engine gains 391 changes from llama.cpp: per chip tuning makes generation faster on every Apple chip from M1 to M5, a memory leak on long conversations is fixed, images are prepared as models expect so every vision model reads better, and video prompts cost about half - Eleven new model families, four small enough for a phone: Granite SWA, Granite Switch 3B, Nanbeige 4.2 3B, and Pocket TTS, plus BailingMoE3, Kimi K3, MiniMax 01, Qwen3.8 Flash Next, Muse Glimmer, dots3-note, and speculative decoding for GLM-4.5-Air - The MLX engine gains 54 changes: models load up to 1.8 times faster, Qwen3.5 and Qwen3 mixture of experts models generate faster, large images no longer exhaust memory on Qwen3 VL, Qwen3.5 VL, and GLM-OCR, and tool calls keep whole numbers and long values intact - New MLX models: Muse Glimmer, 30 billion parameters, reads images and calls tools, Helium and TranslateGemma translate, and Jina Reranker v3 and BGE sort results - An email run report of ten or more messages gains a search box and a filter over senders, subjects, summaries, categories, and folders, matching accented names and full width Chinese - What the email assistant created and learned now sits directly under the summary, so undoing an action no longer means scrolling past every thread - A document outline tells VoiceOver how deep each entry sits and where it jumps, and the reader's loading indicators are named and report progress BUG FIXES - Web pages in a chat or agent workspace opened with a server not found error instead of the page, images and all - A document whose first heading came partway through lost everything above it, and a # inside a code block was read as a heading. Both are fixed, and code blocks stay whole - The search boxes for choosing a model, finding a server, and picking from a list stayed in English whatever the app language Open Doc Scanner, point it at a PDF or a photo, and watch a page become text without leaving your device.

FEATURES - Agent Works toward one goal over many turns. It plans, calls tools, keeps its own files, and delivers a report you can share. - Skills A folder of instructions and scripts that teaches the AI how you want a job done, and becomes a tool it calls by name. Import one from a link, a zip, or a folder, or write it in the app. - Text To Audio Turn text, PDF, ePub, or Word into one audio file with on-device speech models. Nothing is uploaded, and a long conversion resumes where it stopped. - Chat About One Email Start a chat about a single thread from its report, answered by your private model. - Alerts On Events And Reminders Events and reminders created from your mail now get an alert. - Cache Details And Statistics Each reply shows how much of the prompt came from the provider's cache and what it saved. IMPROVEMENTS - Health data tool removed: get_health_data and the Apple Health integration are gone, and no Apple Health data is read - Cached input is billed at the cheaper rate, so costs and spending limits reflect real spend - Engine updates: llama.cpp b10333 adds a GPU math fix and new models, MLX 20260810 reads long prompts 9 percent faster - Python gains numpy, Pillow, and regex on iPhone and iPad, draws charts, and scripts can import others beside them - Saving, listing, and deleting files are now three named tools, so the AI picks the right one first time - Every tool's description and input rules were audited, so tools are chosen and called correctly - Agent runs read their plan back from their own folder and carry reasoning forward - Skills written for desktop tools work here, with command line steps translated - Act on an email while reading it, write your own reply, long inbox runs resume after iOS pauses them, and sender rules and report lists collapse older months - A blocked or verification page is recognized from the server's response, not the wording - Permanent deletes ask you to type DELETE instead of the item's exact name - Statistics gains cache columns and charts, an About page, and VoiceOver chart descriptions BUG FIXES - With iCloud Drive off, the app leaves iCloud alone, and clearing leftovers is yours to start in Settings - Chats with automatic tool selection no longer fail before starting on some providers - MLX models including Gemma 4, Granite 4, and LFM2.5 generate replies again - Pocket TTS packs are repaired automatically, and Kokoro voices in six more languages work - Shell commands chained with &&, ||, and ; run every step, and output is complete with no internal tracing - Flight status, delay, gate, terminal, baggage, and airport boards work again - Subtitle downloads no longer come back empty on a language mismatch - Sub-agents run on on-device models, freeing the previous model first - A rethink keeps tool steps under the previous reply, and a failed retry keeps that reply - The File TTS tab is now called File ASR, which is what it does - In Statistics, Refresh reloads the current tab, cards match height, and headings follow your language Thanks to everyone who reported the File ASR name, the Kokoro voices, and the Pocket TTS packs. Open Agent, give it a goal, and see what it comes back with.

FEATURES - Email Assistant A new mini app that triages your inbox. Connect any IMAP account with an app password, tap Process Inbox, and the AI reads unread mail, flags what needs a reply, and files the rest into folders under an AI label. Choose to summarize, organize, or also move obvious junk. Each run makes a dated report where every decision undoes with one tap, and your corrections teach your preferences. Passwords stay in the Keychain, content is never stored, and nothing is deleted. Mail can also become calendar events, reminders, and notes. - Name Your Pipeline Steps Give each step in an AI Keyboard pipeline its own name, so it shows your label instead of "Custom". - Reusable Saved Steps Save any named step to a personal library and reuse it across pipelines, each an independent copy. - Video on Local Vision Models MLX vision models can now understand video, not just images. Attach a short clip; set the frame count, rate, and size per model. - TurboQuant for Long Chats An opt-in option that keeps your whole conversation in on-device memory by compressing cached values, so long chats use less memory with a small quality trade-off. - Import MLX Models from ZIP Add your own MLX model from a ZIP on your device, no download needed. Tap Import From Local, pick a zip with the model folder, and it installs. IMPROVEMENTS - Local Engine Update (llama.cpp b10091) Two-bit (Q2_0) models now run on the GPU on Apple Silicon, and Qwen3-VL reads layouts more accurately. Adds Laguna code models and completes DeepSeek V4 support. - MLX Engine Update Major memory savings for vision models: Qwen3.5 VL and Qwen3 VL handle long and multi-image prompts with far less memory, and multi-turn image chats stay consistent. Generation also stops cleanly when the app backgrounds. - Engine-Aware Model Settings Settings now show only the options the model's engine actually uses, each tagged with its engine. MLX models now respect Top-K and Min-P. BUG FIXES - Memory Profile Stays On Topic The memory profile now appears only when memory is on and a sector is selected, and no longer blends unrelated sectors together. - Dismiss The Suggestion Bar The word suggestion bar can now be closed with an X, and no longer covers the pipeline picker. - No More Duplicate Imported Models Fixed a model imported from a local GGUF file reappearing as extra "Recovered from" copies after relaunch. Earlier duplicates are cleaned up. - Local Chat No Longer Fails When Memory Is On Local models with memory enabled no longer show a false error. With both models on-device, the chat keeps running and Apple Intelligence saves memories. - Local Models Remember Your Conversation Small on-device models no longer claim they cannot remember earlier messages. A hidden bookkeeping line fed to the model as an instruction is gone. - Long Chat Setting Stays Put Changing how a local model handles conversations that outgrow its context window now sticks, instead of reverting when set inside a chat.

FEATURES - Open Keyboard On Microphone The AI Keyboard can open straight into dictation, ready to record the moment it appears. - Auto-Process After Dictation Stop recording and your pipeline runs automatically, returning the result in place with one-tap Undo. - Dictation Sound Cues Subtle sounds mark when recording starts, when processing begins, and when your text is ready. - Live Dictation Preview Your transcription previews live as you speak and is inserted the moment you stop. - Multi-Language Keyboard Type across all 11 app languages, with Chinese Pinyin and Japanese candidates, Korean Hangul, and European accents. - Inline AI Instructions Tap AI and type an instruction like "intro to Beijing"; it is filled in place while the rest of your text stays put. - Memory For The AI Keyboard Attach a memory sector so Process and inline AI use your identity, preferences, and style. Read on device, off until you opt in. - Knowledge Wiki Turn your documents into a personal, cross-linked knowledge base the AI builds and maintains, stored on device as plain Markdown. - Add Sources From Anywhere Build a wiki from PDF, Word, EPUB, web pages, Markdown, text, CSV, photos, videos, and links. - Ask Your Knowledge Base Ask in plain language and get an answer from your sources, linked to the exact pages. Choose Quick or Thorough. - Interactive Knowledge Graph See how everything connects in a live map: pinch, drag, and tap any point to open its page. - Automatic Overview and Insights Each wiki keeps a living Overview the AI rewrites as you add material, plus AI Graph Insights into the biggest themes. - Wikis In Your Language Pick a language and the AI writes every page and answer in it, across all supported app languages. - Git Sync For Your Wiki Back up and sync any wiki with GitHub or another Git service: publish, pull, clone, and restore full history. - Wiki Health Check A one-tap check finds pages that link nowhere, broken links, and missing pages, with an optional deeper AI review. IMPROVEMENTS - One Tap To Polish After Dictation After you dictate, VoiceOver focus moves to the Process button and reads it aloud, so polishing is one tap. - Keyboard Steps Keep Your Language Built-in steps like Rewrite, Casual, and Summarize now stay in your text's language instead of switching to English. - Rewrite Your Full Selection Process now rewrites the exact text you select, or Select All for the whole field, handling longer text reliably. - Local Engine Update (llama.cpp b9993) New model families including Hunyuan v3, DeepSeek V4, MiniCPM 5, and Unlimited-OCR, faster speculative decoding, and stronger model-file security. - MLX Engine Update (mlx-swift-lm) Gemma 3 prompts up to 2.6x faster on Apple Silicon, more reliable Qwen3/Qwen3.5/Qwen2.5 VL vision, plus Mixtral and Mamba2. BUG FIXES - Keeps Your Audio In High Quality While Dictating Keyboard recording no longer drops Bluetooth, AirPods, or the speaker to phone-call quality.

FEATURES - AI Keyboard A new system-wide keyboard brings Privacy AI into any app where you type. Turn it on once, then in Mail, Messages, Notes, or anywhere else, tap the mic to dictate by voice, or tap Process to rewrite, polish, translate, or reshape what you have written, replaced in place with one-tap Undo. Build your own pipelines (ordered steps such as Polish then Translate) or start from ready-made presets like Proofread, Email, and Summary. Everything runs through the app on your device, so your text stays private and never leaves your phone. - Image Prompt Wizard Building a good image prompt is now much easier. Tap Prompt Wizard on the image generation screen to browse a curated library of prompt templates by their sample picture, pick one, and fill in its options. You can also type a single subject and have AI fill in every field for you in one tap, then preview the finished prompt before you generate. Each template credits its original author. IMPROVEMENTS - Local Engine Update (llama.cpp b9754) The on-device inference engine was updated to the latest build for faster and steadier local models. Multimodal models process attached images more quickly, large models now show real progress while they load, and more models support speculative decoding for quicker replies. Updated graphics and audio components improve performance and stability across the board. - MLX Engine Update (mlx-swift) The Apple MLX inference engine was refreshed for lower memory use and better stability. Vision models now read long prompts in smaller pieces, sharply cutting peak memory so large image chats run more reliably on device. Speculative decoding automatically steps aside when memory is tight, so it never slows you down. This update also adds the Gemma 4 12B Unified and Nemotron text-diffusion models and fixes loading of Qwen3.5, Qwen3-Next, and nomic-embed models.

FEATURES - Smart Chat Automatic Model Routing Smart Chat picks the best AI model for every message automatically, so you never have to choose. Set up your model tiers once (a fast model for simple questions, a capable model for complex work, a reasoning model for hard problems, and a local model for anything private), then just type. Smart Chat reads each message for complexity, privacy sensitivity, and content type and routes it, with a badge on each reply showing which model answered and why. Personal or sensitive messages stay on a local on-device model, you can set per-setup spending budgets, and you can override any single choice. - Video Input For Local Vision Models Attach a video to a chat with a local vision model (such as Qwen-VL or Gemma) and it describes what happens across the clip. The app samples evenly spaced frames and feeds them to the on-device model, so it sees motion and multiple moments rather than a single still. This matches cloud models, fully on device and private. IMPROVEMENTS - Faster Sharing To Chat Sharing a file such as a video now opens the chat right away and prepares the file inside it, with a progress overlay, before sending automatically. No more waiting on a loading screen. If the file cannot be processed, the chat still opens and explains what went wrong. Works in regular chats and Smart Chats. - Resume Interrupted Model Downloads Large model downloads now pick up where they left off instead of restarting from zero. If your connection drops, you cancel and restart, or you leave the app and return, the download continues from where it reached, saving time and data on multi-gigabyte files. Applies to all model types. - Resend Keeps Both Replies Resending now keeps the original AI reply and adds the new one as a variant instead of replacing it. An inline picker (such as 2 of 3) appears on the latest reply so you can flip between versions and pick one. Switching a variant also updates what the assistant carries forward. Works in main chat and Smart Chats. - Local Engine Update (llama.cpp b9663) The on-device inference engine was updated to the latest build, adding video understanding for multimodal models, faster image processing when several pictures are attached, new model architectures, and audio and graphics fixes. - MLX Engine Update (Gemma 4 and Vision) The Apple Silicon MLX engine was updated for faster, steadier on-device models. Gemma 4 gains optional speculative decoding for quicker replies, and its image understanding uses less memory so long prompts with pictures no longer fail on iPhone and iPad. Qwen vision uses less memory too, and Gemma 4, Falcon H1, and LFM2 are more accurate. BUG FIXES - MiniMax-M3 Image and Video Input Fixed multimodal input for MiniMax-M3. Images no longer fail with an "invalid image detail" error, and attached videos are now sent in a format the model recognizes, so it describes the video instead of ignoring it. Videos over 50 MB show a note suggesting the convert-to-text option.

FEATURES - Sign in with ChatGPT Use your ChatGPT Plus or Pro subscription instead of an OpenAI API key to route Codex models. xAI Grok OAuth also ships. - Live Performance Stats In Chats While streaming, a line shows elapsed time, decode speed, and context usage; an expandable row adds time to first token and token counts. - Exa Web Search Provider Exa joins Tavily, Serper, Brave, and LangSearch: neural search returning full page text, 1,000 free searches per month. - User Profile From Your Memories Each memory sector's Profile button asks an AI model to summarize its long-term memories into a short paragraph that prefixes every chat there. - One-Tap Memory Consolidation A Consolidate button finds near-duplicate long-term memories and merges each cluster into one entry. - Memory Maintenance In One Tap Profile and Consolidate first categorize Uncategorized memories and promote short-term notes to long-term. - Memory Audit History A new Audit History view lists every add, update, and delete with keyword search. - On-Device Model Benchmark A Benchmark button runs a llama-bench style test reporting speed and peak memory, then suggests optimal context, batch, and thread settings. IMPROVEMENTS - Text Selection In Fullscreen Chat Selecting text in the fullscreen chat input no longer slides the main menu open, so editing on iPad stays precise. - Choose Transcription Language When converting video or audio to text, pick the spoken language for every speech engine, or leave it on Auto. - Smoother Transcription Progress Audio-to-text progress now advances steadily instead of jumping backward. - Smarter Duplicate Detection Multiple memories from one turn are evaluated in a single AI call, and near-identical entries are skipped. - Memory Search Knows Profile From Facts The cached user profile is sent separately from matching memory rows for a cleaner prompt. - llama.cpp Engine b9553 Updated from b9279 to b9553 (269 changes), adding DeepSeek-OCR 2, Granite 4 Vision, Gemma 4 vision and audio, and DeepSeek V3.2 sparse attention. - MLX-Swift LM Engine tag-20260607 Faster, more accurate on-device vision: corrected Qwen2.5-VL attention, SmolVLM2 up to 9x faster, plus audio input and more reliable tool calling. - Sharper Camera Focus for Documents View Assistant locks focus more reliably on close-up subjects, and you can tap the live preview to focus. BUG FIXES - Soul Greeting Now Appears From the Model View Starting a chat with a Soul from model settings now shows its configured first message. - Attaching a Video to Chat Fixed Video Processing Error: file doesn't exist when attaching a video. - Empty Transcript From iOS Speech Fixed the iOS speech recognizer returning nothing on a clear recording. - Convert to Text Could Hang Fixed the iOS Speech Analyzer getting stuck at Analyzing audio; it now falls back to the system recognizer. - Legacy Memories Now Reachable Memories with an Uncategorized chip were invisible to Consolidate and Profile; all maintenance flows now pick them up. - Memory AI Calls No Longer Stall On First Event Fixed Profile, Consolidate, and Categorize calls returning empty when the first AI event was misread as the end. - Memory List Refreshes After Maintenance After Profile or Consolidate, Uncategorized chips on just-categorized rows disappear immediately. - VoiceOver Input Stays Reachable After Sending Fixed the chat input collapsing after sending; with VoiceOver it stays visible and focusable.

FEATURES - Video Understanding From Any Link Paste a link and the assistant pulls page text, title, and captions. Missing captions trigger on-device transcription. Social links merge post text with attached-video captions, and links with long tracking parameters no longer truncate. Works with most major video and social platforms. - Real-Time Flight Search New search_flight tool returns status, delays, gates, and baggage with fallback across FlightAware, AirLabs, AviationStack, AeroDataBox. IMPROVEMENTS - Faster first reply in a new chat (skips the relevance scorer). - Keyboard opens automatically in empty chats. - Input bar no longer collapses while a draft is in the field. - Quick Ask, LLM Council, Siri, sub-agents no longer disturb an open OpenAI Responses chat. - llama.cpp b9279: HunyuanOCR/HunyuanVL unified, DeepSeek-OCR resize fix, faster Metal, NvFP4 LM head, sliding-window crash fix. - MLX-Swift tag-20260522: Gemma 4 ~24 percent faster, three Gemma 4 crashes fixed, Idefics3 prefill in one pass, Qwen3.5/Qwen3VL/GlmOcr M-RoPE stable. - GGUF tool calling: Gemma 4, Qwen, LFM2 reliably invoke tools via OpenAI function-calling schema. - Faster first reply from local GGUF: Metal kernels pre-compile at load. - Novel Writer on small local models: outline parser recovers from missing tags, character names stay stable, stray markers no longer leak in. - Banner on small local models showing context size and output cap. - Reasoning models (GLM-5.1, DeepSeek, OpenAI o-series) now get budget for hidden reasoning plus full structured output. - Chapter correction pass extends chapters under 70 percent or condenses over 130 percent. - Token Usage charts now count thinking tokens for reasoning models. - Duplicate bottom chat action bar removed; all actions in the top-right menu. BUG FIXES - Web search falls through Tavily, Brave, LangSearch when first returns nothing. - macOS screens stuck in light-mode colors now follow Dark Mode immediately. - Clear Chat on OpenAI Responses drops the server reference and wipes the per-chat workspace. - Fixed Gemma4-E2B/E4B crash on 32 KB Metal threadgroup devices (incl. iPhone 16 Pro Max). - Fixed GGUF metadata read crash on external drives or iCloud I/O errors. - Fixed MLX long-prompt crash on iOS; prompts fed in 512-token chunks on iOS/Mac Catalyst. - Fixed MLX vision models ignoring an image after a text-only reply. - Fixed duplicate "Recovered From..." entries on cold launch. - Custom assistant generation uses a fresh random seed (no more identical results). - Verbose VoiceOver reads the full response instead of only "Response complete". - Collapsed thinking preview no longer reads stray letters under VoiceOver. - Bullet markers decorative; VoiceOver reads item text directly. - LaTeX/math blocks expose a "Math expression" label for VoiceOver.

FEATURES - Quick Text Home-Screen Shortcut: Long-press app icon for a fullscreen AI editor; Translate, Polish, Expand, Concise in one tap. - Civitai Image Generation in Media Studio: SDXL, Flux, and Pony checkpoints with img2img, inpaint, and LoRA support. - Advanced Local Model Tuning: KV quantization, batch size, RoPE, and YaRN controls in GGUF model settings. - Chapter Illustrations and Cover Generation: Auto-generates character sheets, chapter art, and covers; embedded in EPUB. - AI Character Generation: Up to 10 AI-built characters from synopsis; add or remove individually. - Character Concept Sheets On Demand: Generate a portrait for any character from its row menu. - Mini-App REST Endpoints: LLM Council and Novel Writer as REST APIs with SSE streaming. IMPROVEMENTS - Smarter Tool Call Deduplication: Duplicate tool calls return a cached result, stopping loops. - Onboarding Shown Only Once: Welcome screen appears exactly once on first install. - Quick Ask Full Chat Interface: Full streaming view with bubbles and in-toolbar model switcher. - Image Preview Pinch to Zoom: Image attachments open with pinch-to-zoom, pan, and rotation. - AI Toolbar in Fullscreen Text Editor: One-tap transforms with side-by-side review panel. - Tappable Links in Tool Errors: Tool errors include deep links to tool settings or iOS permissions. - llama.cpp b9090: ggml v0.11.0, KleidiAI v1.24.0; MiMo v2.5, MiniCPM-V 4.6, Sarvam MoE, Granite 4.0. - MLX-Swift LM Engine tag-20260509: Fixes Qwen3.5 VL crash; tests expanded to 18 families. - Character Relationship Web: Role/sentiment nodes with chapter-by-chapter relationship timeline. - Mood Arc and Thread Timeline: Mood strip and thread timeline with plant/reinforce/payoff markers. - Clearer New Project Status: Novel Writer projects show "Ready" instead of "Configuring". - mlx-audio-swift tag-20260509: MOSS-TTS (Qwen3), Silero VAD v5/v6; fixes Qwen3-ASR loops. - FluidAudio tag-20260509: Kokoro ANE adds Mandarin G2P; fixes Metal crash and diarization tail. BUG FIXES - Media Studio About View: Fixed wrong about screen in Media Studio. - Chat Input Menu on iPad: Fixed menus hidden behind keyboard. - Mistral Tool Call Compatibility: Fixed 422 error on Mistral tool calls. - Lock Screen Activity Leak: Fixed stuck Live Activity after force-quit. - Repeated iCloud Data Found Prompt: Fixed false prompt for empty iCloud directories. - Image Editor Crash on Close: Fixed crash closing image editor mid-drawing. - Remote Model Title Display: Fixed provider-only titles for manually entered models. - Stability AI Endpoint Routing: Fixed 400 error for stable-image-core/ultra. - Stability SD3/SD3.5 Model Names: Fixed 404 from wrong API names (catalog vs v2beta). - HuggingFace GGUF Download Import: Fixed "Invalid Model File" (local path vs remote URL). - Duplicate Chapter Title Prefix: Fixed doubled titles like "Chapter 1: Chapter 1: ...". - FluidAudio Startup Crash: Fixed main-thread crash scanning FluidAudio models at launch.

BUG FIXES - Fatal Launch Crash on CoreData Migration Failure Fixed a fatal launch crash that could terminate the app immediately at startup on devices where the local database had been opened with iCloud Drive previously enabled and later disabled. The persistent store would be forced into read-only mode, automatic schema migration would fail, and any subsequent save would crash. The app now always enables persistent history tracking, retries store loading with iCloud disabled if the first attempt fails, and guards every database write so a missing store can never trigger an uncatchable exception. Affected users on 2.0.14 will recover automatically on the next launch with this build.

FEATURES - Background AI Generation Remote AI models keep generating when you switch apps or lock your screen. A Live Activity tracks progress and notifies you on completion. - Notification Preferences Toggle completion alerts for AI responses, mini-app tasks, and model downloads in Settings > Notifications. - iCloud Leftover Data Cleanup When iCloud Drive is off, "Manage iCloud Leftovers" removes leftover app data from a previous install. - Completion Notifications Sent when a council session, novel chapter, or health report finishes. - MOSS-TTS-Nano-100M Voice Synthesis Compact bilingual Chinese/English on-device TTS with voice cloning from a short audio sample. - Unified Voice Selection in TTS Preview TTS test panel now has Preset, Cloned, and Design tabs for all local TTS models. - PocketTTS Voice Cloning Voice cloning for all six language packs via the Cloned tab in the TTS test panel. - Image Generation Media Studio generates AI images: pick a provider, describe your prompt, and generate. IMPROVEMENTS - Mini-App Background Generation Council, Health, and Novel Writer keep running with a Live Activity and Cancel on the Lock Screen. - Clearer iCloud Migration Instructions Sync buttons replaced with step-by-step migration instructions in Storage Settings. - Storage Deletion Category Breakdown Deletion confirmation lists each category with file count and size. - Live Stage 0 Role Assignment Council's role-assignment step streams in real time with a thinking disclosure for reasoning models. - Full Language Support Council, Novel Writer, and Health prompts now display in your app language. - Three New Game Theory Scenarios Three new scenarios added: Ultimatum Game, Trust Game, and Public Goods Game. - Multilingual Novel Import Chapter heading detection now covers all 11 supported app languages. - Per-Section Report Regeneration Each Health analysis section can be re-run independently. - Local AI Engine Update (b9008) Faster IQ-series kernels on Apple Silicon, Nemotron Nano 3 Omni GGUF, Qwen3/LLaMA fix. - MLX Audio Engine Update (20260502) MOSS-TTS-Nano added; faster TTS/ASR stop; quicker Qwen3-ASR streaming cancellation. - FluidAudio Engine Update (20260502) Cohere Transcribe chunks audio over 35 s; PocketTTS adds 9 language packs; Parakeet TDT v3 uses int4. BUG FIXES - SVG Image Crash in Markdown Fixed a crash when opening SVG images in a markdown chat file. - Share Sheet Chat List Refresh Chats created via the iOS Share Sheet now appear in the list immediately. - Back Button Hidden in Stage Manager Fixed the chat back button obscured by Stage Manager controls on iPad. - Local Tool Parameter Parsing Fixed tools silently failing when parameter names arrived in camelCase. - Fresh Install Showed Incorrect iCloud Status Fixed "Syncing data from iCloud" on launch and false checkmarks when iCloud Drive was disabled. - Backup Import Preserved the Wrong iCloud Setting Fixed the importer enabling iCloud Drive when restoring from an iCloud-on backup.

FEATURES - Share to Privacy AI: Share text, links, images, PDFs, EPUBs, Office docs, audio, or video from any app via the iOS Share Sheet. - Python Code Execution: AI writes and runs Python 3.14 on-device with numpy, Pillow, httpx, openpyxl, and cryptography. - GitHub Search: Search repos, code, issues, and users; fetch README and release notes. Anonymous (60 req/hr) or with a PAT (5,000 req/hr, full code search). - Noise Reduction: DeepFilterNet on MLX filters background noise before on-device ASR for cleaner transcripts. - Qwen3 TTS Voice Cloning: Reference speaker encoded once and reused across the full response. IMPROVEMENTS - Chat Input Bar: Unified card layout, pill send button, single expandable menu for attachments and settings. - Context Compaction: Older turns summarized automatically when chats grow long; tap Compact at any time. - iCloud Drive Sync: Disabling saves preferences locally; re-enabling resumes cloud sync. - Model Picker: Rows show context window, token limits, and tool-call support. - Icon Button Reliability: Buttons respond across their full frame; fixes tap-but-no-action on iOS 26. - Novel Plot Consistency: STATE.md append-only with AI compaction; FACT.md and summaries use hybrid tail/relevance selection; full outline in every chapter prompt. - Persistent Review Model: Novel Review model choice saved across sessions. - Per-Device Writer Model: Writer model saved per device, not synced via iCloud. - MLX Audio Upgrade: DeepFilterNet V1/V2/V3 (100x real-time); Parakeet batch TDT decoder with bfloat16. - FluidAudio Upgrade (tag-20260425): Diarizer separates speakers (97% F1); iOS 26.4 heap crash fixed; no Cyrillic in English; Kokoro expands abbreviations; Cohere Transcribe added (14 languages, Apple Neural Engine). - LLaMA Upgrade (b8933): Metal matrix-multiply optimized; Reka Edge 2603 vision support; M-RoPE fix; CVE-2026-21869 patched. - Health Reports: 9 new templates, 20+ new data types, per-section data scoping. BUG FIXES - EPUB Extraction: Multi-file TOC chapters now extract all linked content, not just the first file. - Last Message Reappears: History saved correctly when the message list is empty. - Deleted Message Stays Visible: Chat view updates immediately after a message is removed. - iCloud on Fresh Install: iCloud folders created only after you explicitly enable iCloud Drive in Settings. - Spurious iCloud Messages: CloudKit monitoring and sync gated on iCloud Drive preference. - Model Not Found with iCloud Off: File search checks both local and iCloud paths; paths with spaces or special characters resolve correctly. MLX path resolver also fixed. - Writer Accessibility: VoiceOver expand toggle fixed; Actions rotor added. Non-CJK default chapter target raised from 6,000 to 10,000 chars. - Chapter Outline Jank: Expanded rows stay full-width while editing title or summary fields. - Diarizer Model Picker: Shows only models compatible with the active engine (MLX, FluidAudio, or batch).

FEATURES - Text to Video Generation: New chat mode; configure aspect ratio, resolution (up to 4K), and duration; generates via OpenRouter; first-frame thumbnail shown inline. - OpenRouter Video Generation: Additional OpenRouter video models with aspect ratio, resolution, duration, audio, and seed options. - Live Activities for Downloads: Lock Screen and Dynamic Island progress for GGUF, MLX, and CoreML downloads; includes Cancel button. - Local Audio VAD and Diarization: Supports VAD and speaker diarizer models for live transcription. - Multiple Custom Scenarios: Save named View Assistant scenarios with prompt and vision settings; existing prompts auto-migrate. - Health Analysis: New mini-app generates AI-powered reports from Apple Health data; benchmarked against ACC/AHA, ADA, NSF norms; includes a workout readiness score and follow-up chat. IMPROVEMENTS - Fullscreen Text Editor: Expand button opens a near-fullscreen view from any text editor. - Smoother Chat Navigation: Faster open; background monitoring no longer triggers reloads. - Background Download Completion: MLX and CoreML managers handle background URL session events. - CSV Table Header Alignment: Uniform height when column names wrap. - MLX Inference Upgraded (tag-20260419): Gemma 4 text-only; fixes Gemma 4 VLM prompts, Qwen3 tool calls, Mamba/hybrid KV cache, BERT crash on long inputs. - MLX Audio Upgraded (tag-20260419): Fixes Kokoro TTS phonemization; caches Qwen3-TTS embeddings; fixes Cohere int8/int4 loading; Qwen3 ASR context decoding. - LLaMA Engine Upgraded (b8763-b8843): Qwen3 audio; fixes crash on tiny images and Metal Flash Attention; M-RoPE decoder for vision. - FluidAudio ASR Upgraded (tag-20260419): Unified TDT for Japanese; fixes dropped short words; minimum audio down to 300ms. - Feedback Auto-Save: Alerts when no email; saves draft as you type; restores on return. - Smarter Memory Recall: Recency ranking, diversity reranking, and progressive compression on all devices. - Small-Context Model Support: Writer handles truncation silently and auto-recovers. - Smooth Chapter Endings: Detects mid-sentence ends and smooths to a natural boundary. - Novel Writer Error Reporting: Pipeline errors shown as alerts. - Detailed Analysis Scene: New View Assistant scene with per-person physical details and strict OCR. BUG FIXES - Crash with Negative Seed: Fixed crash with a negative model seed value. - Crash Checking Speech Models: Fixed startup crash checking Whisper availability. - Tool Calls on Ternary-Bonsai MLX: Fixed dropped tool calls caused by empty-string vocabulary token for the opening tag. - Gemma 4 Quantized KV Cache Crash: Fixed by routing to updateQuantized() with dequantize. - MLX Memory Not Released: Metal buffer pool released on unload or backgrounding; fixes "Not enough free memory" errors. - Novel Writer Chapter Blank: Fixed blank display after loading; prevents reload on sheet dismiss. - Novel Writer About View: Fixed wrong content; now shows Novel Writer info.

Urgent update for macOS: this release fixes a crash that prevented Privacy AI from launching on macOS (iPad app on Mac). BUG FIXES - macOS First Launch Crash on Initial Settings Screen Fixed a crash that occurred immediately on first launch for users running Privacy AI on macOS (iPad app on Mac). The initial privacy settings screen failed to load. The app now launches and displays the first-run setup screen correctly. FEATURES - Food Search with Full Nutrient Profiles A new search_food tool queries the USDA FoodData Central database and returns complete nutrient data inline for every result, including macros, vitamins, minerals, and fiber. No follow-up request is needed to see the full breakdown. You can request up to 25 results per search. The tool works out of the box using a DEMO_KEY; if you need higher rate limits, you can enter your own FoodData Central API key in the tool settings. - Exercise Search by Name, Muscle, or Category A new search_exercise tool searches the wger.de open exercise database. You can find exercises by name, target muscle group, or category. When your query matches a known muscle or category, the tool runs parallel lookups to return richer results including exercise descriptions, targeted muscles, and required equipment. No API key is required. IMPROVEMENTS - Health Analysis Improved The get_health_data tool now covers 27 additional data types across six new categories.

FEATURES - Initial Settings: analytics, iCloud sync, passcode choices on first launch. - Reddit Search: posts/comments by keyword, subreddit, sort, and time range. - Novel Writer Review: AI evaluates novel across 10 literary dimensions. - Gemma 4 On-Device: gemma-4-e2b-it-4bit via MLX on Apple Silicon. - Docs.OCR Model: on-device document OCR via llama.cpp. - Metal Toggle: enable/disable Metal GPU for GGUF models from Basic settings. - Context Overflow Policy: Rolling Window, Compact, Truncate Middle, Stop at Limit. - Backup and Restore: .zip or encrypted .pabackup; selective restore. IMPROVEMENTS - Novel Writer: synopsis for all types; editable writing rules; edit settings; chapter context menu. - View Assistant Scenarios: 6 built-in; Full Scene Analysis; inline prompt on iPhone. - View Assistant Accessibility: audio ping, voice dictation interrupt (iOS), Magic Tap. - View Assistant Quality: matches device language, strips markdown, no preambles, UI polish. - mlx-audio-swift 20260412: Kokoro iOS crash; Qwen3 TTS 0.6B (7 speakers/10 langs); Cohere Transcribe. - FluidAudio: 2.2-2.8x faster batch transcription; Nemotron 80/160ms streaming ASR. - llama.cpp b8763: Dots.OCR, Gemma 4 Audio, Step3-VL-10B; Q1_0 Metal complete; MERaLiON-2. - MLX Swift 20260412: speculative decoding, parallel tool calls, Swift 6 fix. - TurboQuant: flash attention KV quantization (Q8_0) to reduce memory. - Protocol Inspector: all llama.cpp/MLX logs; zero overhead when disabled. - AI Memory Dedup: two-phase AI-evaluated detection; in-place fact updates. - Backup: import confirmation dialog; atomic imports that roll back on failure. - Edit Council In Place: edit all settings from the chat view. - HuggingFace Import: numbered steps, MMPROJ support, confirmation dialogs. - Device RAM View: GPU Memory, CPU, thermal; Resource Monitor toggle here. - Inference HUD: compact layout, minimize to thermometer icon, thermal color. - VoiceOver: debug log icons marked as decorative. BUG FIXES - Statistics: cost breakdown totals now correct. - Three crashes fixed: keyboard sheet conflict, background thread UI, LazyVStack. - Novel Writer: XML tag remnants removed from chapter titles and EPUB. - View Assistant: camera session fully resets on restart. - TTS language picker shows full locale names (no more duplicates). - TTS voice picker resets on model switch. - Memory dedup race condition fixed. - Memory import count reflects only saved operations. - Backup import is atomic; failed imports roll back cleanly. - Backup password encryption strengthened. - Date settings now survive backup round-trips. - Backup password mismatch error readable by VoiceOver. - Council sessions reload full message history on return. - Chat messages no longer vanish during tool call replies. - Chat: blank-screen bug and scroll overshoot fixed. - Navigation menu: VoiceOver background isolation fixed. - Navigation menu: observer leak fixed. - Navigation menu: first-render animation crash fixed.

FEATURES: - Device RAM Management: New side menu page shows live RAM stats, loaded model list by type, memory tips, and an "Unload All" button with haptic and toast feedback. - Real-Time Streaming ASR: FluidAudio and MLX Audio engines now show a live transcript panel with partial hypotheses during recording; non-streaming engines (WhisperKit) are unchanged. - Gemma 4 Support: Local inference supports Gemma 4 and Gemma 4V (vision) with full tool-calling on E2B/E4B models. - Bonsai-8B 1-Bit Quantization: PrismML Q1_0 1-bit quantized models run on-device with full Metal acceleration. IMPROVEMENTS: - Navigation bar titles added to MLX Audio and FluidAudio model list views. - EXIF metadata (camera, date, lens, GPS) from attached photos is now sent to the AI. - iPad chat query editor stays open while scrolling; keyboard no longer dismisses on scroll. - Tool sheet shows a "Manual | Auto" toggle with an explanation card in Auto mode. - Small context models (4K-8K) get compact tool lists; large context models (64K+) get richer descriptions. - Local model default context scales with file size: under 1GB = 20k tokens, 1-2GB = 14k, larger = 8k. - Swipe-to-open-menu gesture active in list views, disabled in settings with sliders. - FluidAudio upgraded to tag-20260406: Japanese and Mandarin ASR models, PocketTTS session prefill, crash fixes. - MLX Audio Swift upgraded to tag-20260403 with MLXFast for Metal GPU acceleration across all STT models. - Embedder no longer loads on every chat start; loads only when past messages or documents exist. - MLX model loading is 10-15x faster with a streamlined tokenizer. - llama.cpp b8690: Q10 quantization, HunyuanOCR, Granite 4.0 chat template, Gemma 4 and KV cache fixes. BUG FIXES: - Voice picker now shown for all TTS models (FluidAudio and MLX Audio) that offer multiple voices. - ASR falls back to the per-engine language setting when no language is explicitly provided. - ASR test view now shows error messages in red instead of dismissing silently. - "Set as Default ASR" now shows a confirmation alert and updates state immediately after confirmation. - Switching ASR engines now correctly clears the previous model's default flag. - Language picker added to the ASR test widget embedded in model detail previews. - Spoken numbers, currencies, and percentages now transcribed as formatted text ("$200", "50%", "3rd"). - MLX Audio TTS models now store voice preferences independently per model. - TTS and ASR re-verified after storage clear; app falls back to system engine when files are missing. - Stopped generation no longer leaks tokens into the next query; in-flight network requests cancelled on stop. - MLX vision models no longer crash with repetition, presence, or frequency penalty settings. - Council chat history now loads correctly after app restart. - Writer chapter text selection no longer crashes when the copy/paste menu appears. - iPad Council Chat input bar now rises above the keyboard when tapped.

FEATURES - View Assistant New mini app in the Apps menu. Uses your device camera and a vision AI model to describe surroundings in real time, reading responses aloud via TTS. Configurable capture interval (1-10s), prompt, and resolution (512x512 to 1024x1024). Designed for blind and low-vision users with full VoiceOver support. - HuggingFace Inference Provider Selection Choose which inference provider powers HuggingFace API calls: hf-inference, nebius, groq, fal-ai, and more. Open API Keys settings, select HuggingFace, and pick a provider. Applies to chat, image, and video generation. Model list filters to show only supported models. - HuggingFace Model Search Search by keyword when importing a HuggingFace model instead of typing the full repo ID. Results are filtered by model type (MLX or GGUF). Import sheets now open full-screen. - Expanded LaTeX Math Support Approximately 125 additional LaTeX commands: 100+ symbols (negated relations, binary operators, Hebrew letters, arrows), Dirac notation (bra, ket, braket), delimiter sizing (big, bigg, Bigg), boldsymbol, and operatorname. Unsupported expressions render as readable monospaced text. IMPROVEMENTS - Unlimited Conversations with Apple Intelligence Long Apple Intelligence conversations no longer end when context fills up. The app summarizes history and continues seamlessly in a fresh context. - Instant Startup When Offline Detects offline state immediately and loads cached config, eliminating a 6-second or longer hang at launch. - Smoother Streaming Markdown Rendering Chat list rebuilt on a stable foundation. AI responses with tables and rich formatting stream in without scroll jumps or layout resets. - AI Gateway Screen Saver Tap "Screen Saver" in Server Status to show a dim overlay with drifting stats. Keeps the screen on during long server sessions. - Storage Settings as Standalone Page Storage settings moved to its own dedicated page in Settings, previously embedded in System Settings. - Novel Writer: Clearer Project Start Options New project now shows three labeled choices at the top: Choose a Template, Import a Novel, Create an Empty Novel. - Novel Writer: Smarter Outline Review AI evaluator now provides concrete fix suggestions alongside each issue. Evaluator and refiner see previous review rounds to avoid repeating failures. Pass threshold lowered from 8.5 to 7.5. - MLX Audio TTS: Language Picker and Generation Parameters New language selector (Auto-detect or specific language). Four new parameters exposed: Top P, Repetition Penalty, Repetition Context, Streaming Interval. - Local Inference Engine Upgrade (llama.cpp b8461 to b8565) 110 upstream commits. Adds DeepSeekOCR vision encoder support. Reasoning API passes reasoning_content across turns for DeepSeek-R1 and QwQ. Metal GPU matmul2d fix and load-failure segfault fixed. BUG FIXES - Soul with local llama.cpp/MLX models no longer causes a Jinja template error. - Models moved between local and iCloud storage no longer disappear from the list after refresh. - Model deletion now removes files from both local and iCloud storage. - Tool settings sync batched to a single DB write at startup, reducing main-thread I/O. - Fixed crash when model list updates during a background file availability check. - Fixed crash when chat scrolls to a new message during a pending layout update. - VoiceOver: side menu hides background content when open; Writer posts status announcements throughout the writing workflow; progress indicators report percentage; disclosure groups announce expanded/collapsed state. - Voice chat no longer crashes when starting recognition while already active. - Health data tool: steps, distance, and flights now summed (not averaged) per time bucket; sleep total duration double-counting fixed. - EPUB export: fixed stray text fragments between words caused by an attribute-cleaning regex applied to all HTML instead of only tag attributes.

Features - Novel Writer Complete AI fiction studio for long-form novels across 20 genres. The AI builds an outline then writes each chapter with genre-tuned style and consistent voice. A Consistency Database tracks characters, world rules, and events to prevent contradictions. Enhanced Writing Rules (20+) cover show-don't-tell, dialogue subtext, and genre-specific guidance. Quality Report flags AI writing tells. Works on a 9B on-device model. Export as EPUB. - Soul: Character Roleplay Studio Turn any model into a fully realized character. A soul defines name, personality, scenario, and background, injected into every chat. Import or export TavernAI character cards (PNG). Supports custom avatar, opening message with alternate greetings, post-history instructions, tags, and a Lorebook (keyword-triggered lore, constant entries, probability, injection positions). Runtime macros resolve dynamically. Status bar shows the character's name with a sparkle animation. - Markdown Rendering Rewritten with accurate LaTeX math rendering and text selection anywhere in markdown, including code blocks. Improvements - Move Local Models Between Device and iCloud Storage Move GGUF and MLX models in both directions: free iCloud quota or sync across devices. - Presence Penalty for Local Models Available for local GGUF and MLX models in Generation Parameters. - Parse Special Tokens Setting for Local Models "Parse Special Tokens" toggle (on by default) controls whether special tokens in GGUF model output are treated as control signals. - Weekly Pro Subscription Plan Weekly Pro plan now available alongside monthly and yearly. - Screen Stays On During AI Tasks Screen stays on during generation and TTS playback to prevent iOS from reclaiming local model memory. - Local Inference Engine Upgrade (llama.cpp b8355 to b8461) 99 commits: Mistral Small 4 support, vision pipeline improvements, prompt corruption and Jinja fixes. ggml 0.9.8. - MLX Inference Engine Upgrade (tag-20260315 to tag-20260321) Fixed tool calling for XML function format models (e.g. Qwen3.5-2B). LFM2 parser handles arguments with parentheses. - MLX Audio TTS Engine Upgrade (tag-20260315 to tag-20260321) Two new voice models: Fish Speech S2 Pro (Fish Audio) with VQGAN tokenizer and FishS1DAC codec. - Sound Effects and Haptic Feedback for AI Events Sound effects for generation, errors, and clipboard. Haptic feedback on macOS via Force Touch. Disableable in Settings. By Enes Deniz. - VoiceOver: Full-App Accessibility Labels Labels added across every screen: chat actions, TTS controls, nav menu, Prompts, Settings, Reader, Audio/ASR, Tools, MCP, Memory, Novel Writer, Soul, Council, and more. - VoiceOver: Navigation Structure and Announcements Section headers and message senders carry the heading trait for rotor navigation. Decorative elements hidden. Loading and download progress announced. - VoiceOver: Verbose Announcements Opt-in toggle in Settings > Accessibility announces generation start and reads the response aloud. Off by default. Bug Fixes - Apple Intelligence Hidden on Unsupported Devices Now hidden on unsupported devices (e.g. iPhone SE 2022). - VoiceOver: Interaction and Focus Fixes Fixed: focus stuck on storage directory entries; menu not dismissible (Z-scrub added); drag handles announced as interactive; expandable sections not reporting state; chat input bar buttons unlabeled. - Council Folder Hidden from Chat Folder List Council folder no longer appears in the chat folder list. - MLX Audio TTS Stability Fixes Fixed a crash under memory pressure and a race condition on new TTS requests. - Calendar and Reminder Tool Stability Fixed startup crash. Now initialize lazily and detect permission revocation without restart.

Features - API Rate and Credit Monitoring Tap the gauge icon after any AI response to check API rate limits and credits. OpenRouter shows credit balance and dollar amounts; other providers show rate-limit data from response headers. - Game Theory Mode in LLM Council A new Game Theory mode has council members negotiate across multiple rounds before the chairman evaluates outcomes. - Mermaid Diagram Rendering Mermaid code blocks now render as native charts (flowcharts, sequence diagrams, Gantt, and more) with a toggle between the chart and raw source. - search_news New Providers search_news now supports four free news providers: NewsData.io, News API, GNews, and Currents. - App Store Connect Management manage_appstore_connect lets you manage App Store Connect through AI chat: apps, reviews, TestFlight, in-app purchases, and sales reports. - omlx Self-Hosted MLX Inference omlx is now a built-in provider for a self-hosted MLX server at localhost:8000/v1. - GZip File Support Privacy AI opens .gz files directly, decompressing them automatically. Improvements - Paper Search Covers Medical Research search_paper now searches PubMed alongside arXiv, and the year filter is fixed. - More Relevant arXiv Results Multi-word searches now require all terms, so "melatonin safety" returns specifically relevant papers. - Fullscreen Camera on iPad Take Photo now opens fullscreen on both iPhone and iPad, replacing the partial sheet that clipped the preview. - llama.cpp Engine Upgrade (b8355) Upgraded from b8185. New PEG-based autoparser improves tool calling reliability. Checkpointing fixed for Qwen3VL, Llama4, and Qwen2VL. NVFP4 quantization added. - MLX Audio Engine Upgrade (20260315) Parakeet V2 ASR improved; Qwen3 TTS faster. Adds SenseVoice, FireRed ASR 2, IBM Granite Speech 4, Echo TTS, and three new audio codecs. - MLX Swift Engine Upgrade (20260315) Qwen3.5 Vision and MoE Vision now supported. Tool call XML parsing fixed. RoPE improved across 25+ model families. GLM-OCR added. - Larger Models on High-RAM Devices Privacy AI requests the increased-memory-limit entitlement on iPad Pro, enabling larger models and bigger context windows. - LM Studio Reasoning Compatibility The Responses API now maps LM Studio's reasoning_text events correctly so reasoning models like Qwen3.5 work as expected. - Statistics Charts Improved Model Usage and Cost charts are now horizontal bar charts with exact value labels. Daily Usage is now a line chart showing Total Cost. - Configurable Message Streaming Chat Theme Settings offers Typewriter, Chunked, or Immediate display with refresh rate 5-60 FPS. - LLM Council Enhancements Queries fetch pasted URLs for context, long queries collapse after 100 characters, and interrupted sessions resume from the last completed stage with history maintained. - Streaming Response Performance Markdown parsing now handles only new characters per update, reducing CPU usage and jank in long replies. - Dark Mode App Icon The app icon switches to a dark variant on iOS 18 with Dark or Automatic icon theme. - SVG Image Sharing SVG files shared to Privacy AI are now converted to PNG for AI vision analysis instead of being dropped. Bug Fixes - Qwen3.5 Tool Calling Fixed Qwen3.5 GGUF and MLX models (0.8B, 2B, 4B) can now call tools correctly. - Long Audio File Transcription MLX Audio (Qwen3-ASR-0.6B) now processes long audio files without errors across all common formats. - File Sharing via Action Extension Binary files (ZIP, PDF, etc.) shared via the Action Extension are now correctly received and attached to chat. - Search Fallback in Auto Mode search_web Auto mode now correctly falls through to the next engine when a provider returns zero results. - Share Sheet Blank on First Tap Fixed the share sheet appearing blank on first open from code, TTS audio, chat, and certificate export. - Appearance Theme Fixed an iOS issue where switching to Light or Dark mode left no way to return to the system appearance. Auto is now available.

Features - MLX Audio Engine On-device audio engine with TTS, STT, VAD, and speech-to-speech models, including Qwen3 TTS 0.6B and Qwen3 ASR 0.6B. Download from the new MLX Audio Models section under Local Models. - Live Transcription & Real Time Translation Continuous streaming transcription with local MLX STT and VAD models. Enable real-time translation as audio or video plays. Fully on-device, no cloud needed. - Internal Model Upgraded to Qwen3-0.8B The built-in model for internal tasks is now Qwen3-0.8B. New Qwen3.5 models (0.8B, 2B, 4B) are available in both GGUF (text + vision) and MLX (text only) formats. - Cost Report Export Export usage statistics as a Markdown report from the Cost Estimator. Share it to another app or copy to clipboard, with a clear Remote/Local breakdown. - AI Gateway Vision Support The local AI Gateway now accepts image inputs via remote URLs or base64 data URIs in the standard OpenAI format. Works with vision models like PaddleOCR-VL-1.5 and GLM-4.6V. Improvements - Cost Report Charts Redesigned Charts replaced with horizontal bar charts showing exact data labels. Capped at top 20 entries. iPad layout reorganized. - Flexible Health Data Aggregation The get_health_data tool accepts an optional aggregation parameter (raw, per-minute, hourly, daily, monthly, quarterly, yearly). Default behavior unchanged. - Subtitle to Article Shared Action Share SRT or VTT subtitle files to convert them into clean articles with corrected transcription and organized sections. - llama.cpp Engine Upgrade (b8145 to b8185) Latest upstream performance and compatibility improvements for local inference. - MLX-Swift Engine Upgrade (20260224 to 20260302) Adds Qwen3.5 and Qwen3.5 MoE support. Prompt cache save/load works correctly. LFM2 RoPE config fix. Wired memory kept stable during streaming on iOS 18+. - MLX Audio Engine Upgrade (20260224 to 20260302) Reduced memory allocations in audio resampling. Downloads detect zero-byte files from interrupted transfers and report clear errors. - New Qwen3.5 Models (2B and 4B) Available in GGUF and MLX model lists with strong multilingual performance and improved reasoning. - Status Bar Visible on iOS The iOS status bar now shows time, battery, and signal indicators. - Voice Preview in MLX Audio Model Details Audition any downloaded TTS model's voice directly from its detail page without changing your global TTS settings. - Set as Default ASR from Model Detail Downloaded STT models show a "Set as Default ASR" button in their detail view. - Live Waveform in ASR and MLX STT Views Recording indicator displays animated amplitude bars driven by real microphone input level. - Whisper Language Preference Persisted Selected language saved to iCloud and applied automatically on subsequent sessions. - Faster Processing Pipeline Large file content is trimmed in a single pass instead of split across multiple rounds, reducing overhead. Bug Fixes - System Prompt Applied Across All AI Models Fixed a gap where chats without an explicit system prompt bypassed the model's configured default. A unified priority chain now resolves the correct prompt for all model types. - Local LLM Crash on Certain Vision Models Fixed a crash with Qwen2-VL and Qwen2.5-VL by detecting incompatible memory layouts and using a safe full-context reset. - Local LLM Premature Context Resets Fixed position tracking that overestimated context usage, causing unnecessary resets mid-conversation. - LLM Council Token Tracking Token usage from Council sessions is now recorded in the Cost Estimator. - File Attachment Preview Fixed in-chat document preview showing raw text instead of rendered markdown. - Download Progress Restored on Re-entry Progress bar now correctly reattaches when returning to a model detail view during an active download. - LLM Council Thinking Model Tool Calls Models like deepseek-reasoner and Gemini-3-Pro now work as Council members with tool calling enabled.

Features - LLM Council Inspired by Andrej Karpathy's llm-council, Privacy AI now supports LLM Council natively. Create a council of AI models (GLM, Minimax, Kimi, local models) that all respond to the same question simultaneously. No external orchestration needed. Use for multi-perspective research, debate, critical thinking, and business strategy analysis. - GLM-5 Auto Tool Selector Support GLM-5 models from Zhipu AI require a strict tool call whitelist. The auto-tool selector now correctly meets GLM-5 requirements, enabling automatic tool invocation just like other supported models. - PaddleOCR-VL Vision Model PaddleOCR-VL-1.5 is now in the recommended vision model list. Purpose-built for OCR — accurately extracts text from images, scanned documents, receipts, and complex layouts. Runs entirely on-device. Improvements - Delete Calendar Events with AI The manage_calendar tool now supports deleting calendar events. Ask AI to remove an event by title; provide a date to narrow the search when multiple events share the same name. - AI Gateway Model Search Filter The AI Gateway model list now includes a search field to filter by model ID instantly. - Copy Model ID from AI Gateway List Long-press any model row in the AI Gateway list and select Copy Model ID to copy it to the clipboard. - Clearer AI Agent Tool Name The sub-agent spawning tool is renamed from manage_task to run_agent, eliminating confusion with task lists and calendar management. Existing workflows are unaffected. - Clearer AI Task Planning Tool Name The workflow tracking tool is renamed from manage_todo to manage_plan, better reflecting its purpose of coordinating structured multi-step plans. Existing workflows are unaffected. - Collapsible Folders Section in Chat Sidebar The Folders section in the chat list sidebar can now be collapsed and expanded, like the Pinned and Recent sections. - Auto Tool Selector Now Works with Local Models The tool list for local models is optimized to ~8,000 tokens, freeing enough context for on-device models to use the auto tool selector. - llama.cpp Engine Upgrade (b8089 to b8145) Upgraded to b8145 with support for PaddleOCR-VL, GLM-OCR, Kanana-2, JAIS-2, and Full Modern BERT. ARM64 q5_K inference is faster, multimodal flash attention is now configurable, and GGUF integer overflow protection is added. - MLX-Swift Engine Upgrade (20260218 to 20260224) Upgraded to tag20260224. mlx-swift moves from 0.30.3 to 0.30.6 with stability improvements, a LoRA adapter initialization bug is fixed, and strict Swift concurrency is enforced in MLXEmbedders. Bug Fixes - Decimal Input in Local Model Settings Fixed: typing a decimal point in temperature, top-p, and repeat penalty fields was immediately erased. These fields now correctly accept decimal input. - AI Gateway Large Request and Stale Settings Fixed two issues: large prompts (4,000+ tokens) caused HTTP 400 errors from incomplete request buffering; stale cached model configs ignored context/token setting changes. Both are now resolved. - Voice Selection Not Taking Effect Fixed a key mismatch in TTS settings save/read paths. Changing voice, speaker, sample rate, chunk size, or speed now correctly applies to speech output. - Thread Count Setting Not Affecting TTS Fixed multiple issues: thread count was never passed to the TTS engine at startup, the engine only initialized once so later changes had no effect, and UI changes weren't saved. The engine now reinitializes with the correct thread count on every change.

Features - Edit File Tool Make precise string replacements in workspace files. AI specifies exact text to find and replace, using surrounding context to ensure uniqueness. - Run Shell Tool Execute shell commands in the workspace directory. On iOS, ~150 commands available including ls, grep, sed, awk, find, tar, curl with full pipe and redirect support. On macOS, full native zsh with all system commands. Sandboxed to the chat directory. - grep_file Tool Professional regex text search in workspace files. Streams line-by-line to handle files up to 10MB. Features: invert match (-v), count-only mode (-c), before/after context (-B/-A), glob patterns for multiple files. - Automatic MLX Model Type Detection Automatically detects whether a downloaded MLX model is an LLM or VLM. No manual configuration needed - download from HuggingFace and the app identifies and loads the correct type. - Tool Selection for Chat with Document A new Select Tools section on the Reader's "Chat with Document" setup screen lets you enable any local or MCP tools the AI can use during the session. Improvements - Thinking View Auto-Scroll: Disabled during streaming when the reasoning view is expanded, so you can read earlier steps without the view jumping to the bottom. - Copy As Plain Text: New context menu option on assistant messages strips markdown formatting (bold, italic, headers, code fences) while preserving line breaks and list markers. - HuggingFace Repository ID Clear Button: Tap X to instantly clear the repo ID field when importing MLX or GGUF models. - MLX-Swift Engine Upgrade (tag-20260203 to tag-20260218): Model weights and KV caches pinned in GPU-wired memory, reducing generation pauses on iOS. New model support: MiniMax, MiMo v2 Flash, LFM2.5-VL. Adds native function-calling and token streaming API. - llama.cpp Engine Upgrade (b7988 to b8089): New model support: Tiny Aya, Nemotron Nano 12B v2 VL (vision), Kimi Linear hybrid. Metal concurrency improvements for stable inference on iOS and macOS. - Enhanced Tool Descriptions: More accurate descriptions and parameter aliases (e.g., "query" also accepts "search", "search_query", "keyword") across all local tools. - Enhanced Tool Call Compatibility: Improved GLM-5 compatibility. Parameter alias system accepts alternative names for more resilient tool calls. - Cloudflare Agent Support: Added 'text/markdown' to all URL readers for full Cloudflare 'Markdown for Agents' specification support. - Multiple Thinking Blocks Merged: Multiple thinking blocks now merged into one unified view. Missing opening tags handled correctly so all reasoning content is captured. - Reliable Tool Calls on Small Models: Tool call parser now recovers when small thinking models (e.g., 3B Nanbeige) omit closing tags, ensuring tool calls are correctly parsed and executed. Bug Fixes - Memory View Restoration: Fixed missing memories not appearing in the Memory view. - AI Gateway Streaming Order: Fixed scrambled text chunks from concurrent Task race conditions in chunk delivery. - File Upload Filename: Fixed .md suffix incorrectly appended to text file uploads. - Startup Crash (Pre-1.7.0): Fixed crash when loading saved chat configs from versions before 1.7.0. Older configs load with sensible defaults. - Tool Call Limit Cloning: maxToolCall setting now preserved when cloning, forking, or branching chats. - OpenRouter Gemini Thinking Text: Fixed truncated responses when thinkingText is present. Bidirectional prefix checking properly separates thinking and response content. - Model Context Setting: Fixed context value above 4096 reverting after confirming the "High Context Warning" alert. - Year-Month Chat Deletion: Fixed group deletion removing chats across all folders. Scoped to current folder only. - Discord Community Link: Updated to a permanent invite link (previous temporary link had expired). - Reader Document Context: Large documents now handled via workspace summary to stay within model context limits.