Features - Rethink Button New "Rethink" lets the AI continue reasoning without re-running tool calls—useful for exploring alternatives or extending analysis while avoiding repeated tool costs. - Min-P Sampling Parameter Support Added "Min-P (Minimum Probability)" for compatible providers (including OpenRouter and DeepSeek). Min-P filters tokens by a minimum probability threshold relative to the most likely token. Higher values (e.g. 0.1) are more focused; lower values (e.g. 0.01) are more diverse. - Interleaved Thinking Support Added "interleaved thinking (reasoning expansion)" for advanced reasoning models across providers, including DeepSeek (deepseek-reasoner), OpenRouter (minimax-m2, deepseek-v3.2, gpt-5.2, claude-opus-4, kimi-k2-thinking), MiniMax (minimax-m2), Kimi/Moonshot (kimi-k2-thinking), and Google Gemini (gemini-3-pro-preview). This preserves reasoning across multi-turn conversations and tool calls for better continuity and planning. - Claude API Interleaved Thinking Support Claude API now support native interleaved thinking for CLAUDE 4+ conversations. - VCard and ICS Support Added support to read and convert ".vcf" and ".ics" into structured Markdown, including contact fields and calendar items (events, tasks, locations, attendees, reminders). Improvements - iOS 26.2 Platform Update Updated build target to "iOS 26.2" for latest platform optimizations, security updates, and improved stability. - Improved Tool Call Display Multi-tool-call messages now auto collapse/expand. If more than 3 tool calls exist, only the latest 3 show by default, with a control to expand and view all. - Consistent Citation Card Citation preview cards in Sources now use a consistent size (160pt × 160pt) for a cleaner layout. - llama.cpp Engine Upgrade (b7332 to b7402) Upgraded local inference engine (+67 commits): optimized Metal backend for iOS Mamba/SSM, improved performance and threadpool stability, improved CLIP handling, added vision model support (Qwen3VL, KimiVL). - Tabbed Remote Model Configuration Remote Model configuration now uses tabs: "Model", "Pricing", and "Settings". - Default Model Setting Moved default model selection from Siri settings to system settings. Updated settings views with a more unified theme. - Chat List Swipe Action Added bottom padding so the floating action button doesn’t block swipe actions on the last chat. - Quick Ask Keyboard Handling Improved keyboard behavior: tap the response area to dismiss, and refined the input field sizing for more consistent interaction. - Redesigned Sharing System Redesigned share extension to support more file types, including "ICS" and "vCard". - Article Retrieval from X Added support to retrieve and convert "X.com" article URLs into clean Markdown. - Enhanced Tool Search with Improved UI Tool search now matches tool names and descriptions. Tool toggles include short, human-friendly usage descriptions. - Markdown Rendering Optimization Optimized streaming Markdown rendering by normalizing list indentation and cleaning formatting during streaming. Bug Fixes - Image Sharing Restored Fixed image sharing to Privacy AI from other apps. - Clear Chat Messages Fixed clearing all messages in a chat. - OpenAI API Compliance Fixed `parallel_tool_call` being sent to OpenAI when no tools were included; it’s now only sent when tools are used. - Markdown Theme Fixed Markdown color scheme picker defaulting to light mode; it now follows the current system appearance. - Reader Context Menu Actions Fixed Reader context menu actions (Copy, Select & Edit, Share, Read Aloud) after the cache migration. - Whisper Model Retry Failure Fixed failures when retrying Whisper downloads by cleaning partial download folders before retry. - Duplicated Model Settings Fixed saving changes to duplicated remote models; updates now persist and show immediately.

Features - ZIP Archive & Excel Support Reader now opens ZIP archives directly and displays all contained files. Excel workbooks with multiple sheets are fully supported - OpenRouter Protocol Support Native OpenRouter integration with support for OpenRouter-specific features. Configure web search via the :online suffix, choose native or Exa search engines, and customize max results and context size from a dedicated settings panel. This extends beyond OpenAI-compatible behavior and unlocks the OpenRouter feature set while preserving full inspector/debug tooling. - Media Discovery Tools Three new zero-setup tools: – Books: Google Books search with NYTimes bestseller rankings. – Podcasts: iTunes/PodcastIndex discovery with episode transcript support. – Games: Steam search for pricing, deals, discounts, and news. - Talk to Document Reader now includes an integrated chat panel. Switch between Read and Talk modes using a segmented control. In Talk mode, choose any AI model and ask questions. The entire document is automatically injected as context, enabling accurate, document-aware responses for research, analysis, and long-form reports. - Quick Actions Long-press the app icon and tap Quick Ask for an instant chat window with streaming responses. Uses the same default model configured for Siri. Ideal for quick, ephemeral questions; no chat history is saved. - Duplicate Model Remote model configurations can now be duplicated with one tap. All fields: API key, URL, system prompt, tools, parameters, are cloned, and the duplicate opens immediately for editing. - Chat Organization Chats are grouped by year and month (e.g., “December 2024”). Each section includes a one-tap bulk-delete control for easier cleanup. Improvements - Collapsible Message Previews Messages collapse to ~5 lines for faster scrolling. Tap Show more to expand. The latest message is always expanded. - llama.cpp Upgrade (b7222 → b7332) Includes critical GPU-residency stability fixes on iOS, preventing Metal memory eviction during long sessions. Adds FP16 Flash Attention overflow fix, Mach-O build corrections, and support for the new Rnj-1 model. - Expanded File Type Support 30+ new formats including SRT/VTT subtitles, XML/YAML, code files with syntax detection, SVG, and FLAC. - Rich Preview Cards for Citation Links Markdown links now render as portrait preview cards showing favicon, domain, and multi-line title with smart caching. - Enhanced Model Pricing View Pricing UI now includes capability metadata (vision, attachments, reasoning, tools, structured output), knowledge cutoff, release dates, open-weights status, and dedicated cache-pricing fields. - Smarter Rating Prompt Logic In-app review triggers now use milestone-based heuristics for better timing. - New Siri Shortcut: “Ask Privacy AI” Quickly query any model directly via Siri or the Shortcuts app. - Improved HTML to Markdown Conversion More accurate heading and list preservation for cleaner document imports. - Table & Image Indexing in Reader Tables and images now appear in Reader’s outline under level-6 headers for easier navigation. - Auto-Saved Drafts Unsent chat drafts are saved automatically and restored when returning to the conversation. Bug Fixes - Memory Import Infinite Loop Resolved an infinite loop caused by reusing the active chat model during memory import. A fresh model instance is now used. - Streaming Stop Control Streaming now halts immediately when Stop is pressed. - Crash on Long Paragraphs Large markdown blocks are now safely segmented to avoid layout overflows. - Message Resend Context Loss System prompts are now correctly restored when resending messages. - Shared Content Ordering When sharing both text and attachments, the app now waits for attachment processing before sending to ensure complete context delivery.

Features - File Processing Mode in Sharing UI When sharing files, you can now choose how the content is handled before sending it to the model. "Send as is" preserves the original file format. "Convert to Markdown" extracts text and converts it into lightweight Markdown to reduce token usage—recommended for most workflows. - HuggingFace API Privacy AI now supports the HuggingFace API. Use your HuggingFace token to access HF models All outgoing requests are captured by the built-in inspector, and the integration works seamlessly with mitmproxy for complete end-to-end API debugging. - Nous Research Provider Nous Research is now available as a new remote provider. Connect directly to their API to access additional cutting-edge models and integrate them into your workflows just like other cloud providers. - Markdown Code Block Syntax Code blocks in Markdown now render with syntax highlighting for 40+ programming languages. AI-generated code is easier to read and automatically adapts to light and dark mode for a consistent viewing experience. - Max Tool Calls Control You can now set a maximum number of tool calls per response in each model’s settings. This gives you precise control over the trade-off between thorough reasoning and cost/performance. Higher limits (10–20) allow deeper research flows with multi-step tool use; lower limits (3–6) prioritize faster and more economical responses. - Network Traffic Inspection via mitmproxy All protocols now support full network traffic inspection through mitmproxy. This makes it easy to diagnose protocol issues, inspect raw request/response payloads, and understand how data flows through each provider. Enable proxy mode under Settings > Proxy. Improvements - llama.cpp Engine Upgrade (b7140 to b7222) This release brings a major performance upgrade to the local inference engine. The Metal backend now supports Flash Attention with head size 48, improving memory efficiency and speed. ARM CPU performance for Q4_K quantized models has been significantly boosted through optimized i8mm and dotprod instructions—resulting in faster inference across all Apple Silicon devices (M1/M2/M3/M4 Macs and iPhone 13+). Added support for the Ministral-3-3B model and improved compatibility with LFM2-VL vision models. The engine now also supports model-embedded sampling parameters, enabling GGUF files to specify their own recommended generation settings for more consistent outputs. - Accurate Content Detection The attachment previewer now detects the true document type after import and conversion. When sharing documents, spreadsheets, audio, video, or subtitle files, the previewer automatically identifies the resulting format (Markdown, plain text, JSON, CSV, etc.) and displays the correct viewer. - Large Markdown File Rendering The Reader now handles large Markdown documents with significantly improved performance. Scrolling is smoother, memory usage is reduced, and multi-chapter books, long articles, and technical documentation load more efficiently. Bug Fixes - Action Extension Content Handling Fixed an issue where the Action Extension failed to process certain shared content types. Screenshots (PNG), images, text snippets, and other file formats now attach reliably when shared to Privacy AI. Previously, sharing screenshots could result in empty attachments. - Attachment Preview Improvements Fixed a UI freeze that could occur when previewing large documents. Text now loads promptly and scroll performance is noticeably smoother, especially for large chat histories or extensive Markdown files. - PDF Outline Position Accuracy Fixed a bug in the PDF-to-Markdown converter where section headers appeared at the bottom of the document instead of their correct positions. The outline extraction logic has been rewritten from a recursive approach to a simpler, more reliable loop structure that respects page boundaries and preserves original hierarchy.

Features - Claude API Protocol Support It now supports the Claude API, enabling direct connections to Claude models as well as Z.ai GLM4.6 and Kimi K2 Thinking. Multimodal inputs (text, images, PDFs), tool calling, and extended-thinking modes are supported. The protocol inspector logs all requests and responses in real time across every protocol, making debugging and verification easy. Works with Anthropic’s endpoint and any provider that implements the Claude API. Existing Z.ai and Kimi Code Plan users can connect using their current API keys. - Max Tokens Parameter You can now set max_tokens for both local and remote models. This option is integrated with model settings and is respected consistently across all protocols. - Multiple Named Endpoints per API Key. API keys can now define multiple named endpoints under the same credential, ideal for providers offering multiple routes - Local ↔ iCloud Sync Tools New sync buttons let you migrate chats and data between local storage and iCloud Drive. Syncing skips existing files, shows real-time per-file progress, and reports detailed results. Two-way sync is supported, and Clean Local / Clean iCloud actions include double-confirmation protection. - Storage Management A new Storage Management screen shows file counts and space usage for each directory. Each category includes a one-tap delete button (with confirmation), and switching between local and iCloud refreshes data instantly. - Function Call Rate Control Remote models now support a Function Call Wait Time (0–30 seconds) to delay consecutive tool calls and avoid rate-limit errors. Status messages appear between executions when delays are active. - Subtitle File Support (SRT & VTT) Sharing .srt and .vtt subtitle files now imports them as plain text for translation, caption analysis, and dialogue extraction. Improvements - Smarter Web Search Ranking Search results now use combined keyword + semantic scoring for more relevant answers. - Protocol Type Safeguard Protocol type can no longer be changed from inside an active chat. A warning explains that protocol switching requires creating a new chat from Remote Model settings. This prevents conflicts between OpenAI Chat, OpenAI Responses, and Claude API protocols. - Language-Aware Context Calculation Context and token estimates now use a unified, language-aware system that adapts to English, Chinese, Japanese, Korean, code, and mixed inputs. Output reservation scales with model size, and all components now rely on the same calculation logic. - llama.cpp Engine Upgrade (b7091 → b7140) ARM64 i8mm optimizations deliver 15–30% faster Q4_K model performance on iPhone 13+ and M1+ devices. This upgrade adds RND1 diffusion model support, improves Metal backend stability on macOS 11+, fixes grammar parsing overflow vulnerabilities, and increases solver accuracy. - Protocol Indicator Badges Model logos now display protocol/engine badges: oai, resp, claude, gguf, or mlx, making model configuration clearer at a glance. - Secure API Key Entry API keys are now hidden by default with a reveal/hide toggle. - macOS Keyboard Shortcut Support Shift+Enter now sends messages on macOS, consistent with Bluetooth keyboards on iPhone/iPad. - Modernized Text Editor & Selector UI The editor has been redesigned with improved Edit/View segmented control, inline search with match navigation, updated toolbars, a relocated Done button, character counters, and animated copy confirmations. Bug Fixes - Long Document Processing Fixed a bug where intermediate chunk responses did not appear during long-document analysis. Chunk messages now render correctly. - Model Lifecycle Stability Improved handling of model recycling and memory recovery to prevent unexpected interruptions during long sessions. - Model Parameter Persistence Context and sampling parameters now sync reliably between model settings and chat sessions, ensuring consistent generation across workflows.

Improvements - llama.cpp Engine Upgrade (b7091) Updated to b7091 with improved Metal shader optimizations, faster argsort operations, and accelerated conv2d for better performance on iOS devices. This release adds support for new architectures such as AfmoeForCausalLM (Afmoe MoE models) and introduces new operators (CUMSUM, SOFTPLUS, TRI, SOLVE_TRI) required by hybrid and advanced reasoning models. iOS builds also benefit from enhanced ARM feature detection and improved numerical stability for more reliable inference. - Sharing to Privacy AI Refined the UI layout to make the Share-to-Privacy-AI interface more compact and easier to use. Added a new Refresh button to reload the model list instantly. - Enhanced Keyboard Management in Model Configuration All text fields and text editors in model configuration screens now include a convenient Done button to dismiss the keyboard. For numeric keyboards without a return key, a smart toolbar appears with a dismiss button. This enhancement applies across all model detail views. - Improved HTML to Markdown Conversion The HTML-to-Markdown engine has been upgraded with significantly better noise filtering, especially tuned for X (Twitter) posts. You can now share X post URLs directly to Privacy AI, and the content will be loaded and converted into clean, readable Markdown for analysis. - Smarter Web Search with Reference Links Web search results now include automatically extracted reference links appended to the end of AI responses, making it easier to verify information and explore original sources. - MCP Marketplace UI Optimization Each MCP server’s card layout has been optimized so descriptions can be fully displayed. On iPad, cards now maintain a consistent height across multi-column layouts. Bug Fixes - Missing Local Models Fixed an issue where downloaded local models were not displayed when sharing content to Privacy AI.

Features - Direct Video Input Support Privacy AI now supports OpenRouter’s native video input protocol. You can send video files directly to supported multimodal models: such as Gemini 2.5 Flash, Flash Lite, Pro, and others, without any manual pre-processing. The app handles encoding and upload automatically, offering a seamless video-to-AI workflow. - Upgraded TTS/ASR Engine sherpa-onnx has been upgraded to 1.12.15, adding support for newer TTS models like MatchTTS and expanding available voice options for higher-quality speech synthesis. Improvements - MLX Engine Upgrade Updated to the latest MLX Swift framework with improved vision-model handling and greater stability. Qwen3-VL models now process images more reliably thanks to refined sanitization. Memory usage is optimized on both iOS and macOS, and the new context API enables more flexible prompt construction for advanced RAG workflows. - llama.cpp Engine Upgrade (b7032) Major performance improvements for Apple Silicon, including Metal 4 Tensor API acceleration on M5-class devices, ARM64 SVE optimizations, and hybrid context shifting for better memory efficiency. This update also includes KV-cache optimizations, async buffer-retention fixes, and enhanced support for A19 devices. Expect 20–40% faster inference on supported hardware with improved GGUF stability. - WhisperKit Audio Transcription Boost WhisperKit now uses automatic compute-unit selection, model prewarming, improved VAD chunking, and smarter decoding heuristics. Short audio less than 15s transcribes 10–20% faster, while long audio (>3 min) uses VAD to reduce memory pressure and improve reliability. - Whisper Model UI Refresh The Whisper Models screen has been redesigned with more model variants and clearer descriptions to help you choose the right model for your workflows. - Improved Whisper Model Downloading Whisper model downloads now run in the background, support resume-on-break, and display detailed progress information for a more reliable setup experience. - Modernized About & Acknowledgements UI The About screen has been redesigned, with all third-party libraries updated to reflect their latest versions and descriptions. Bug Fixes - Partial Responses Fixed a bug that caused truncated responses for some models in the Gemini family in OpenRouter. - Audio Transcription Fixed an issue where audio and video files stored in iCloud Drive could not be read correctly on iPhone or iPad during transcription.

Features - OpenAI Responses API Support Privacy AI now supports the OpenAI Responses API, OpenAI’s next-generation protocol enabling stateless multi-turn conversations without resending full chat history on each request. The API also introduces conversation objects for automatic context management, prompt templates with variable substitution, and native prompt caching for faster, cheaper repeated queries. It’s fully compatible with OpenAI and any third-party provider implementing the Responses API standard (including LM Studio). The built-in Protocol Inspector displays every request, response, and SSE stream in real time—ideal for debugging and verifying custom endpoints. - MCP Marketplace (Pro) A new marketplace powered by [MCPRouter](https://mcprouter.co/) lets you browse and install pre-configured MCP servers in one tap. Each integration—automation, image generation, time zones, data analysis, and more—is auto-configured with endpoints, headers, and authentication ready to go. No manual setup required. - llama.cpp Engine Update (b6962) Major CPU optimization for iOS and macOS with ARM64 chunking and Flash Attention improvements. Adds support for vision models like Qwen3-VL, CogVLM, and Janus Pro, offering better multimodal reasoning with reduced memory use on mobile devices. - Interleaved Thinking Support New support for interleaved thinking in models such as Minimax M2, Claude Sonnet 4, and OpenAI GPT-5 (Thinking series). When enabled, the model preserves internal reasoning across turns, enabling self-reflection, planning, and retry logic for complex multi-step tasks. This brings truly agentic behavior—turning one-shot answers into autonomous workflows. - Multi-Image Selection Select up to 10 photos simultaneously from the picker. All selected images are processed sequentially and attached automatically, streamlining multi-image analysis or comparison. Improvements - Simplified MCP Authorization MCP settings now include a dedicated *Authorization* toggle with a secure API-key field. Privacy AI formats Bearer tokens automatically, eliminating manual header composition and reducing connection errors. - Improved Protocol Diagnostics HTTP 404/429 errors now display clearer explanations—showing whether a model is rate-limited or temporarily unavailable—and include actionable recovery guidance. - File Analysis UX New real-time progress indicators during OCR or document parsing give clear feedback through every stage of file analysis. - Collapsible “Similar Models” View The Similar Models section now collapses by default for cleaner layout and can be expanded instantly to compare prices and alternatives. Bug Fixes - macOS Stability Fixed a crash when opening Model Settings on Mac; the Model Detail view now loads reliably across all devices. - Local Model Cleanup Resolved a memory leak caused by failed GGUF or MLX model loads. Failed retries no longer leave stale allocations or cause crashes. - Super Siri Integration Restored full Siri compatibility across MLX, GGUF, Apple Intelligence, and all remote providers. - Dynamic Model Pricing Model pricing now refreshes instantly whenever you switch models in the Model View.

Features - New Vercel Provider We’ve added Vercel (vercel.com) as a new official provider. Vercel offers some of the most cost-efficient AI models available today. Free-tier users receive $5 in credits every 30 days for use with AI Gateway models — a great way to explore their ecosystem at no cost. All Vercel model pricing data has been fully refreshed within Privacy AI. - MLX Vision Models Now Support Tool Calling Qwen3-VL, Gemma 3, Qwen 2 VL, Qwen 2.5 VL, and SmolVLM 2 now support tool calling. These vision-language models can act as text-to-text engines, enabling seamless integration with Privacy AI’s local tools and workflows. Improvements - Enhanced GGUF Model Downloads The GGUF download flow has been aligned with the MLX model process for a consistent user experience. You can now view both the primary GGUF model file and its optional CLIP component (if available). Added Verify and Refresh options ensure all required files are complete before initiating a chat. - llama.cpp Engine Upgrade (b6871) Upgraded the internal llama.cpp engine from b6778 to b6871. 1. KV-cache memory alignment adjustments: removes over-padding and aligns context/buffer ordering, significantly reducing wasted RAM on mobile-class devices. 2. New model support: adds compatibility for LightOnOCR-1B and BailingMoeV2, expanding local model choices.

Features - Memory Support Introduces the long-awaited "Memory" feature. It lets you add, edit, and delete memories that persist across all models. Your AI can now retain personal context even when you switch models. Memory is free for local models and available under the "Pro Plan" for cloud models. All data are stored securely on-device and can optionally sync via iCloud for seamless continuity—high privacy, full control. Improvements - llama.cpp Upgrade (b6804) This major upgrade delivers substantial performance gains on iOS and macOS through optimized Metal backends and ARM NEON acceleration. It adds Flash Attention F32 support, non-padded KV cache for better memory efficiency, and faster normalization operations. Also includes new math ops, optimizer APIs, and fixes for memory leaks and build issues. Expect faster inference, lower memory use, broader model compatibility—including IBM Granite Hybrid—and improved overall stability for on-device AI. - MLX Engine Update Updated to support the latest "Qwen3 VL" and "Gemma 3" models for multimodal and text-generation tasks. - MLX Download Verification MLX models often include multiple component files. The model-detail view now offers a "Verify" button to confirm all files are correctly downloaded; once verified, the Create Chat button becomes available automatically. Bug Fixes - DeepSeek Stability Fixed an issue where chatting with DeepSeek models could cause unexpected crashes. - MLX Model Validation Fixed a crash that could occur when attempting to load incomplete or corrupted MLX models.

Improvements - MLX Engine Upgrade Upgraded the MLX engine to build `8775112`, adding support for Falcon H1, Qwen3 embeddings, Granite Hybrid MoE, and LiquidAI LFM2MoE models for broader local compatibility and faster inference on Apple Silicon. - Updated Recommended MLX Models Replaced the Granite 4.0 H Tiny (7B) model with the lighter Granite 4.0 H Micro model—optimized for iPhones where 7B models may exceed memory limits. Added LFM2 2.6B MLX to the recommended list for improved multimodal performance. - API Detail View Enhancements The API configuration view now displays all available endpoints, letting you see exactly which URLs are used to connect to each remote server. Bug Fixes - Hugging Face Repository ID Resolution Fixed an issue where Privacy AI failed to correctly extract the Hugging Face repository ID from certain downloaded MLX models. - Chat Creation During Downloads Fixed a bug where already-downloaded models could not be used to start new chats while another model was still downloading.

Features - MLX Model Integration Privacy AI now supports the MLX model engine, enabling both text and vision models to run locally. You can directly download models from Hugging Face by entering a repository ID and access token (if required). The new download manager adds resume-on-failure, background downloading, and a model integrity verifier for reliable large-model transfers. MLX models now fully support local tool calls and MCP tool calls, just like GGUF and remote API models. Best of all, MLX model support is included in the Free Plan. - Clipboard and Drag-and-Drop Enhancements The Chat and Reader editors now allow pasting images, videos, and files directly from the clipboard. On iPad, you can also drag and drop files or images into the chat editor — a smoother, more natural way to attach media. Improvements - llama.cpp Upgrade Updated llama.cpp from b6558 → b6692, adding compatibility for new models such as Qwen3 Reranker and LiquidAI LFM2-2.6B, plus broader quantization format support. - Text Rendering Refinement Improved text layout and rendering during “thinking” mode for better readability and consistency across devices. - Smarter Model Switching Switching local models now filters to show only models of the same type — MLX, GGUF, or Apple Intelligence — for a cleaner experience. Bug Fixes - Thinking Text Display Fixed an issue where “thinking” models occasionally displayed incomplete or cut-off lines when collapsed. - Thinking Mode Persistence Fixed a bug where the local model’s thinking-mode setting was not properly saved or applied to new chat sessions. - Single Local Model Enforcement Resolved a critical issue where MLX and GGUF models could run simultaneously, potentially causing memory overflow.

Improvements - GLM-4.6 Integration Added support for the latest GLM-4.6 model under the Z.ai provider, giving users access to improved performance and reasoning capabilities. - IBM Model Branding Introduced official IBM logos for the Granite series and other IBM models, making it easier to visually identify them in the model list. - Enhanced iPad Experience Improved the remote model configuration UI on iPad with a cleaner layout and better usability for larger screens. Bug Fixes - Remote Parameter Saving Fixed an issue where key parameters—including `temp`, `seed`, `top_p`, `repeat_penalty`, `presence_penalty`, `frequence_penalty`, and `reasoningEffort`—were not being saved correctly for remote models. These settings are now preserved as expected. 1. Fix the bug that parameters including "temp", "seed", "top_p", "repeat_penalty", "presence_penalty", "frequence_penalty" and "reasoningEffort" can not be saved for remote models

Improvements - Refined switch model behavior The switch model button now switches between models from the same provider instead of toggling between local and remote APIs. Bug Fixes - Consistent model picker display Fixed an issue where the model picker in chat did not consistently show both remote and local models. - Removed duplicate switch button Eliminated the extra switch model button in the chat UI. Users can now rely on the **Clone** and **Fork Chat** features for a better way to continue a conversation with another model.

Features: 1. Free Plan is live We’re excited to announce a big upgrade to the app with the introduction of the Free Plan. Most of the core features will now be permanently free to use, including: - Local Models and Apple Foundation Model - Natural language conversations - Reader - iCloud sync - Smart web & document-to-Markdown conversion - Audio and video transcription - Full integration with Siri and Shortcuts - All 25 built-in tools (Search, News, Stocks, Weather, Health, Email, Calendar, and more) - Conversation cloning and branching - Advanced export options - Built-in HTML rendering Only a few advanced features: such as "Cloud models", "MCP", and custom API providers, require a Pro Plan subscription. 2. Control the frequency UI refresh rate of text rendering on screen under throttle. For very fast AI models, the fast text rendering to Markdown on screen will make the UI lag and consume a lot of battery. 3. llama.cpp upgrading Upgrade 'llama.cpp' engine version to b6558. Support Liquid AI series models now 4. Model settings copy When clone or fork a chat, if you select same model as current model, all current model's properties like temperature, top_p, top_k, context are all copied to new chat too. Improvements 1. Optimize the type-writer effect. During a chat when AI is typing. Use advanced dynamically throttle and cache to control the UI refresh rate. Make the screen updating smoothly and keep battery life longer. 2. Enable auto-scroll by default. In chat UI, when a new query is submitted and AI starts to reply, the reply text will automatically scroll when next text is coming. 3. According to open-source community suggestions, set the AI's default temperature to 0.3 (from 0.6) and Top P to 1.0 (from 0.9). These are the best experience values for open source models. You can adjust them manually of course. 4. Apple on-device Foundation model supports the streaming Bugfixes: 1. Apple on-device Foundation model is mistakenly treated as a cloud model. In current version, the on-device model is categorized to local model 2. Fix the bug that local model or imported model can not change its title

Feature - Added Apple’s on device Foundation model Added Apple’s on device Foundation model to the local model list. Integrated with 25 built-in tools (search, email, news, and more) and extended support for external MCP tool calling. Improvement - Customizable think tags Users can now define their own start and end tags for self-hosted or cloud models. Bug Fix - Resolved system prompt issue Fixed a bug where system prompts for self-hosted and cloud models could not be saved.

Features - Support for the latest iOS 26 UI and integration with Apple’s on-device Foundation Model. - Introduced a new hands-free “Natural Talk” UI, letting you talk naturally with any model. - Display the “What’s New” screen automatically after each app update. - Rebuilt history handling to support binary attachments. Images, PDFs, and other files are now embedded directly in chat history, so you can reference them in later requests. - Attachment previews now support Microsoft Office documents, Apple iWork files, PDFs, videos, and audio formats. - Added support for OpenAI’s **gpt-image-1** model and HuggingFace’s **black-forest-labs/FLUX.1-dev** image generation model. - Added support for the OpenAI protocol TTS endpoint, including **gpt-4o-tts** and **gpt-4o-mini-tts**, to generate audio from text. - Introduced a brand-new preview tool that can open and export PDFs, images, Office 365 documents, iWork files, videos, and audio generated by AI. Improvements - Raised the minimum OS requirement to iOS 18.6. We will continue aligning APIs with the current and previous iOS versions. - Prevent starting a remote chat session without selecting a model first. - Improved binary file handling (e.g., PDFs) for maximum compatibility with OpenAI-compatible servers such as OpenRouter. Raw PDFs can now be uploaded directly for processing. - In Protocol Inspector, capped displayed text to 2,000 characters to prevent UI freezes with large Base64-encoded requests. - Enhanced history token management to ensure requests and history stay within model token limits. - Added API pricing data for TTS, image generation, and transcription models, including hybrid pricing (per minute, per image, etc.). - Improved chat UI for tool calls, with tool call data now saved and reloaded with chats. Bug Fixes - Fixed an issue where the mini player bar stopped playback when scrolling through chat. - Fixed a bug where refreshing an image query ignored updated parameters such as count and size.

New Features - Super Siri Tool Control – Configure tools when invoking Siri. Enable web search or calendar management directly by voice. - Smarter Local TTS – Pause and resume local text-to-speech from the current position. - Upgraded Speech Engine – Updated sherpa-onnx to 1.12.11 for better pause handling at punctuation. - Large File Processing – Oversized files are now split into chunks with preserved context, enabling summaries, translations, and more at any scale. Improvements - Mini TTS Player – A new toolbar player replaces the blocking dialog, letting you read while listening. It also highlights the current text being spoken. - Multi-File Import – Import multiple files into a single chat for richer workflows. - Smarter Image Text Extraction – When enabled, extracted text is added as the alt attribute in Markdown image links, improving clarity for AI. Bug Fixes - Markdown attachments now render in the Markdown viewer instead of plain text. - URL submissions now extract images according to Reader settings (previously all images were included). - Fixed X.com link handling: sharing or opening X posts in Reader now extracts the full post content.

Features 1. Multiple Attachments in Chats Upload and process several files of different types in a single conversation. 2. Process Multiple URLs at Once Paste multiple links in one message—Privacy AI fetches, merges, and analyzes them together. 3. Latest Excel Support Convert the newest Office 365 XLSX files into clean Markdown in Reader or chats. 4. Smarter Reader Settings New “JS Check Interval” and “JS Idle Threshold” options improve accuracy when converting web pages. 5. Interactive AI Images Rotate, flip, zoom in or out on AI-generated images directly in the app. 6. New Image Editor Draw or edit images with Apple Pencil or touch, then send them straight to AI. 7. Edit Sent Messages Update messages after sending—attachments are preserved automatically. 8. Context Usage Indicator See live token usage in chats, based on server reports or Privacy AI’s estimator. Improvements 1. Friendly Error Messages Errors are now explained in plain, human-readable text. 2. Upgraded Inspector Export full logs or copy partial logs for faster debugging. 3. Smarter Remote Error Handling Server errors are shown as simple explanations, not technical stack traces. 4. Better Message Editing Attachments stay linked when you edit a message. 5. Optimized Token Allocation New algorithm preserves more history while keeping safe space for replies. 6. Improved File Upload Dialog Clear mode descriptions explain technical details, costs, and best use cases. 7. Provider Memory New remote models default to your last used provider, not Ollama. 8. Transparent Subscription UI Free trial details are now explained clearly. 9. Faster Siri & Shortcuts Streamlined stateless workflow improves performance. Bug Fixes • Fixed Siri and Shortcuts not connecting to AI models when the app wasn’t running in the background.

Features 1. OpenRouter Image Generation: Generate images directly through OpenRouter. Try the free Gemini 2.5 Flash Image Preview model today. It also supports OpenRouter ZDR policy 2. Smarter File Import for Media: When importing videos or images, a new dialog lets you configure dimensions. The app automatically scales the shorter side or uses iOS APIs to extract text, saving tokens and costs. You can also choose to send files unchanged to remote AI. 3. Flexible File Handling: For non-media files, you can upload them as-is or convert them to Markdown locally. Markdown conversion reduces cost but may trade off some accuracy. 4. iCloud Model Sync: Download a GGUF model once and use it across all your devices with the same iCloud account. Models load automatically in the background—no need for multiple downloads. 5. Reader Enhancements: Added support for capturing photos directly in the reader and selecting processing methods before importing documents, images, or videos. The reader UI has also been refreshed. 6. Expanded Self-Hosted Server Support: Now works with llama.cpp, vLLM, LocalAI, and Jan AI. 7. HTML in Code Blocks: Code blocks containing HTML can now be rendered and captured as screenshots for easy sharing. 8. Rich Chat History: Chat messages now preserve attachments and generated media, creating a more complete conversation record. 9. Feature Overview: Added a section that lists all features and highlights of the app for easier discovery. 10. Improved Model Pricing Display – Added new fields and improved the UI of model price information shown in the Remote Service and API Keys views. Upgrade - Updated llama.cpp to b6301 for improved performance and compatibility. - Remote Services now remember the collapsed/expanded status for each category. Bug Fixes - Fixed an issue where newly installed apps sometimes failed to display local AI models in the selection list. - Fixed an issue where the application failed to refresh the local cache after remote configuration changes.

Features 1. OpenRouter Image Generation: Generate images directly through OpenRouter. Try the free Gemini 2.5 Flash Image Preview model today. 2. Smarter File Import for Media: When importing videos or images, a new dialog lets you configure dimensions. The app automatically scales the shorter side or uses iOS APIs to extract text, saving tokens and costs. You can also choose to send files unchanged to remote AI. 3. Flexible File Handling: For non-media files, you can upload them as-is or convert them to Markdown locally. Markdown conversion reduces cost but may trade off some accuracy. 4. iCloud Model Sync: Download a GGUF model once and use it across all your devices with the same iCloud account. Models load automatically in the background—no need for multiple downloads. 5. Reader Enhancements: Added support for capturing photos directly in the reader and selecting processing methods before importing documents, images, or videos. The reader UI has also been refreshed. 6. Expanded Self-Hosted Server Support: Now works with llama.cpp, vLLM, LocalAI, and Jan AI. 7. HTML in Code Blocks: Code blocks containing HTML can now be rendered and captured as screenshots for easy sharing. 8. Rich Chat History: Chat messages now preserve attachments and generated media, creating a more complete conversation record. 9. Feature Overview: Added a section that lists all features and highlights of the app for easier discovery. Upgrade - Updated llama.cpp to b6301 for improved performance and compatibility. - Remote Services now remember the collapsed/expanded status for each category. Bug Fixes - Fixed an issue where newly installed apps sometimes failed to display local AI models in the selection list.

- Feature: Local vision models on device. You can now run vision-capable GGUF models locally via our llama.cpp mtmb integration (for example, Qwen2.5-VL 3B Instruct). Add images to a chat without sending data to the cloud. - Feature: Add our first vision model, “Qwen2.5-VL 3B Instruct.” - Feature: Add new suggested models. Added “gemma-3-270m” to the recommended list - Improvement: Faster and more reliable downloads. Local model downloads are quicker and more stable; we also fixed an issue that could start extra download workers when the app moved between background and foreground. - Improvement: Clearer remote model pricing. The price panel now shows “Cache Read (per 1M tokens)” and “Cache Write (per 1M tokens)” so you can estimate costs more accurately when KV cache is used. - Improvement: Easier ways to contact us. In Feedback, you can open your default email app or DM our X account. There’s also an optional email field if you want a reply. Because the app has no user accounts, we can only respond if you leave contact info. - Bugfix: Model list refresh. Newly downloaded local models now appear immediately when you switch models inside a chat.

Version 1.1.23 - llama.cpp Upgrade – Updated from b5950 to b6131 for better GPU performance on Apple devices, more efficient KV cache handling, and support for new models: GLM-4.5, SmallThinker, Qwen3-Embedding, and Hunyuan Dense. - Faster Large Model Imports – Refactored the GGUF file processor to handle 4B+ models with the latest llama.cpp engine, importing quickly without memory overflows. - Perplexity API Update – Added support for the latest models: sonar, sonar-pro, sonar-deep-research, sonar-reasoning, and sonar-reasoning-pro. Model names can now be edited in Perplexity API settings. (Tip: If you already use the Perplexity API, remove it from the API list and restart the app to apply changes.) - TTS Settings Enhancement – The TTS settings view now displays your device’s CPU core count to help you choose the optimal thread count. - Clearer X.com API Key Guide – Added more detailed instructions in Settings for using your X.com API key. - Tool Description Improvements – Updated the search_contact local tool description so AI knows when and how to use it. Related tools like send_email and send_sms now rely on it for recipient lookup. For example: “Send my wife an SMS and tell her I love her” will first trigger search_contact to find the recipient, then create the message. - Better HuggingFace Downloads – Model downloads now run in the background and can resume after interruptions. A prompt will notify you when the download is complete—ideal for large models like Qwen3-4B-Thinking/Instruct-2507. - Tool Search Bar – Added a search bar for local and MCP tools to quickly find what you need as the tool list grows. - UI Tips Added – Added quick usage tips for API Key Management, Tools, and Remote Services to help new users get started faster. Bug Fixes - Fixed: Search bar in Remote Services now filters models correctly. - Fixed: Removed the “Add” menu from Local Tools, since adding external tools is no longer supported. - Fixed: Resolved a crash when sending SMS with the send_sms tool.

1. Offline Text-to-Speech An offline TTS model Kokoros-82M (53 distinct voice styles) is now bundled directly in the app. Any AI reply or Reader article can be spoken aloud entirely on-device, so nothing is sent to the cloud and there are no per-character fees. You can also export the generated audio (M4A,WAV,AIFF) to Files, AirDrop, or any media player for later listening. 2. API Server Templates and Cloning You can duplicate an existing server profile—such as the HuggingFace—to create custom endpoints in seconds. All models, headers, tokens, and endpoint settings are copied automatically. You only adjust what differs (base URL, model path, etc.). This is ideal when a provider exposes multiple endpoints under a single API key or when you run several private vLLM clusters. 3. Built-in GitHub Provider GitHub has been added to the list of internal API provider. 4. Flexible Model Selection for Forked or Cloned Chats When you fork an existing conversation or clone it into a new thread, you can now change the underlying model—local or remote—before the next message is generated. The original chat remains intact, and the new branch inherits the full context and tool settings while letting you compare answers or continue with a faster or cheaper engine. Bug Fixes • The file-import button that disappeared inside chats is back. • Manually entered model names now save correctly in API Key configuration screen.

1. Built-in API Access to z.ai - We’ve embedded native support for https://api.z.ai/api/paas/v4/. You can now call Z.AI services directly from Privacy AI with your token. 2. Perplexity Model Deprecation - The outdated r1-1776 model from Perplexity has been removed. Please switch to sonar-reasoning or sonar-reasoning-pro for continued access via OpenRouter. 3. Expanded OpenRouter Protocol Compatibility - Improved protocol handling ensures better performance and compatibility with the latest OpenRouter models and backends. 4. Major Siri Integration Upgrade - Local Models Now Work with Siri: You can now trigger local models using Siri voice commands. This is made possible by enhanced performance of llama.cpp-based models. - Faster AI Replies for Siri: Adjusted prompt logic helps AI respond within Siri’s strict time limits (typically under 8 seconds). For best results, use fast-response models—avoid long-thinking agents. 5. Improved Thinking Text View - When thinking models produce long outputs, the text now scrolls smoothly with a visible scrollbar—no more UI lockups on lengthy thoughts. 6. Search Tool Now Has Speed/Balance/Quality Modes - Choose your preferred mode for the searching_tool. “Speed” mode delivers up to 3× faster results, perfect for quick lookups. 7. Code Block UI Stability - Markdown rendering has been optimized to remain responsive even when AI outputs include very large code blocks (1000+ lines). No more freezing or slowdowns in the chat UI.

1. Fixed API Endpoint Save Bug Resolved an issue where custom API base URLs were not being saved properly in the API settings screen. 2. Improved URL Input Experience Disabled automatic capitalization for URL fields to prevent input errors when entering API endpoints. 3. Streamlined API Server Creation You can now create a new API server directly from the API Key detail view—making it faster to configure your remote models. 4. Enhanced Launch Feedback Added detailed progress indicators during app startup to show what’s being synced or initialized.