Privacy AI release notes what's new

General

• bug fix
  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.

Local Models

• improvement
  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.

• improvement
  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 fix
  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.

• bug fix
  Retrying Audio and Image Downloads

  Retrying an MLX audio or image model download no longer deletes a download that had already finished.

Local Audio

• bug fix
  Partial Whisper Downloads

  A Whisper model that only partly downloaded no longer shows as Downloaded. It now offers to continue the download instead.

iOS 27

• feature
  Models That Run On The Neural Engine

  Privacy AI can now run models on the Neural Engine, a fourth on-device engine beside llama.cpp, MLX and Apple Intelligence. It is the only one that keeps writing after you leave the app: a long answer carries on while you look something up elsewhere, where a model on the graphics chip stops the moment Privacy AI is no longer on screen. On an iPhone 16 Pro Max a 0.6B model generated 60.3 tokens a second in the background against 76.5 in the foreground, and the same holds inside the AI Keyboard, so these models can answer anywhere you type. Five are available on iOS 27 and macOS 27: Qwen3 0.6B, MiniCPM5 1B, MiniCPM5 2B, Lucy and SmolLM3 3B, with context windows from 5,120 to 16,384 tokens. Each is built for your particular chip before first use, which takes about a minute when a ready-made build exists for your device and around ten minutes when it does not, and the result is kept so you do not wait again. They come in two kinds, though, and only the Neural Engine kind carries on in the background. Privacy AI now reads the model and tells you which you have, when you browse a repository, when you import a file and on its settings screen, and one built for the graphics chip stops cleanly when you leave the app instead of stalling, and is not offered to the AI Keyboard, which could never have used it.

• feature
  A Bigger Apple Model

  Privacy AI can now use Apple's larger cloud model, Private Cloud Compute, alongside the smaller one that runs on your device. There is no account to create and no API key to enter. Apple processes each request without storing it, and your device verifies the server before sending anything. A new Apple Models screen lists both, saying plainly which runs on your device and which runs on Apple's servers, and the cloud one also appears under Remote Models when you pick a model. It is free to use, within the allowance Apple gives your account, and the app tells you when that allowance is running low. Requires iOS 27 or macOS 27 on a device that supports Apple Intelligence.

• feature
  Apple Models Can See Your Pictures

  Attach a photo to a chat using Apple Intelligence or Private Cloud Compute and the model now looks at the picture itself. Until now the picture was quietly left out and those two answered from your words alone, while every other model in the app received it. Ask what is in a photograph, what a chart is showing or what a screenshot says, and the answer comes from the image. Up to four pictures can travel with one message. Once the reply has arrived, any writing found in the picture takes its place in the conversation's memory, so later questions do not spend your context reading the same image over and over. Requires iOS 27 or macOS 27 on a device that supports Apple Intelligence.

• feature
  Models That Cannot See Get A Description

  Reading a picture used to mean reading the words inside it and nothing else. A photograph with nothing written on it came back empty, so a model without vision of its own had nothing at all to work with, and attaching one to a chat simply showed an error. Privacy AI now writes a short description of the picture on your device and hands that over instead. Any model can use it: a small model running on your phone, an older one, or a provider's text-only model that would otherwise charge you for vision. Ask what is in a photo, what a diagram shows or what is happening in a screenshot, and the model answers from the description. It applies wherever the app reads a picture, including an attachment you pick with Extract Text, an image file you open, and the AI reading one for itself. Writing found in the picture always wins, so documents and screenshots behave exactly as before and cost nothing extra. Requires iOS 27 or macOS 27 on a device that supports Apple Intelligence.

• feature
  Core AI Models From HuggingFace

  Core AI models can now be brought in from a HuggingFace repository, alongside the existing import from a zip file, the way llama.cpp and MLX models already could. Because a Core AI repository holds one folder per device rather than a flat set of files, the repository is read first and only the builds this iPhone or Mac can actually run are offered, each already matched to the right one. A repository that holds no Core AI model says so instead of starting a download that cannot work.

• feature
  Speech Recognition On The Neural Engine

  A new on-device speech recogniser runs on the Neural Engine, and it is the only one that keeps transcribing after you leave the app, so a long recording carries on while you do something else. It covers 25 European languages, and the list of languages now comes from the model itself, so one it cannot write is no longer offered. It needs iOS 27 or macOS 27 and about 1.3 GB.

• feature
  28 Kokoro Voices In Half The Space

  A second Kokoro voice pack, built in Apple's Core AI format, brings 28 American and British voices in 756 MB, where the existing pack is 1.4 GB for 54 voices. Reading aloud pauses when you leave the app and carries on when you come back. It needs iOS 27 or macOS 27 on a real device, and it is offered only where it can run.

General

• feature
  One Tool To Read Any File

  Reading a slide deck, a spreadsheet, a PDF, an e-book, a chart or a screenshot is now one thing the AI does, called Extract Text, instead of two separate tools it had to choose between. It works on a file in the chat as well as on a web address, so a chart on a page can be read without saving it first. A spreadsheet or an archive comes back as a list of its parts so a single sheet can be asked for by name. Handing it a recording now gets a clear answer pointing at transcription, rather than a several minute wait inside a tool meant for documents.

• feature
  Now Available In Swedish

  Privacy AI is now fully available in Swedish. Every screen, menu and setting is translated, the AI keyboard has a Swedish layout with word prediction, and the assistant can be told to answer in Swedish. Dictation understands Swedish, and reading a reply aloud uses your device's Swedish voices. Choose it under Settings, App Language, or leave that on System Default and the app follows whatever language your device is set to.

• improvement
  Captions First When Transcribing

  Transcribe Media now looks for the video's own published captions before recognising anything, and only listens to the audio when there are none. Published captions are exact and arrive in seconds. The result says which of the two it used, so a quote can be attributed honestly, and there is a setting to ask for captions only when waiting several minutes is not acceptable. It also accepts a recording you attached or the app created, which previously nothing could transcribe.

• improvement
  Faster Reading Of Text Inside Pictures

  When reading the text inside a web page's images is turned on, the pictures are now read at the same time instead of one after another. A page with nine images could previously hold everything up for the better part of a minute.

• improvement
  Scripts Get Full Tool Results

  When a script calls one of the app's own tools, such as looking up prices, flights or news, it used to receive only a small fraction of the answer, far less than the AI itself gets from the very same tool. Scripts now receive the whole result in almost every case. If a result is still too large to hand over directly, the complete copy is saved into your workspace and the script is told exactly where to find it.

• improvement
  Home Screen On Launch

  Privacy AI now opens on a Home screen: a grid of twelve tiles covering every conversation surface and mini app, from All Chats, Agent and Smart Chats to Doc Scanner, Novel Writer and the AI Keyboard. Reaching a mini app used to take four steps, opening the side menu, tapping Apps, scrolling, then tapping the one you wanted, and each is now a single tap from launch. Tiles can be reordered, pinned into their own section at the top, or hidden away, they line up in even rows, and one that needs a subscription says so on the tile. Everything on Home stays in the side menu as well, so nothing you used to reach there has moved out from under you. On the first launch after updating, the one time tidy up of your files now finishes before the welcome and What's New screens, which could otherwise be skipped.

• improvement
  One Thinking Tool Instead Of Two

  Sequential Thinking has been retired. Today's models think a problem through on their own, and in testing they mostly ignored it and answered without it, while its instructions took up a slice of every request. Critical Thinking remains and now stands on its own, and the space this frees goes to your conversation.

• improvement
  Reading The Text Inside A Page's Pictures

  When the AI fetches a web page it can now be asked to read the text inside the pictures on it, such as the figures in a chart. It stays off unless asked for, because reading pictures takes time, and asking for it fetches the page again rather than handing back an earlier copy that has none.

• improvement
  Chats Kept In iCloud Drive

  When your chats are stored in iCloud Drive, your device keeps only what you have opened recently and leaves the rest in the cloud. Opening an older chat or a mini app project used to freeze the app while it fetched those files, for a long time on a slow connection and indefinitely with no connection at all. Privacy AI now checks whether a file is on the device before reading it, so your lists open straight away. A chat that is still arriving shows a download bar with a real percentage instead of a screen you cannot use, and one whose files are still in the cloud stays in the list marked as not downloaded, rather than appearing to have been deleted. Nothing is ever written over a file that could not be read, so a conversation cannot be replaced by an empty one, and a backup now tells you when files are still in the cloud instead of quietly leaving them out. Opening the app and scanning your models are both faster as well, because neither waits on files it does not need.

• bug fix
  Projects And Chats No Longer Look Deleted

  On a second device, where iCloud Drive keeps most of your files in the cloud, Novel Writer could show "No Writing Projects" with every project still there, and the same went for Knowledge Wiki, Text to Audio, Doc Scanner, LLM Council, Email Assistant and your agents. A project whose settings file had not been fetched was skipped, and nothing ever asked for it again. Those projects now stay in the list marked as not downloaded, and the app quietly fetches them so the rows fill in on their own. Pinning one can no longer replace the real project with the placeholder. Lists also show a brief spinner while they load instead of flashing an empty screen. Separately, the download bar when opening a chat from iCloud Drive reached the end and stayed there, because the app was reading a stale answer to the question of whether the file had arrived. It now tracks the download and clears itself, Cancel opens the chat straight away with what is already on the device, and a download that fails says so instead of opening an empty conversation.

• bug fix
  Search Results Point At The Right Line

  Searching a file reported the wrong line number whenever there was a blank line above the match, because blank lines were not being counted. On a document with paragraphs the number could be out by a dozen lines or more, so following a search result led somewhere else. Line numbers now match what you see, and they agree with the line numbers used everywhere else in the app.

• bug fix
  File Names With Spaces

  Asking the AI to open a file whose name contains a space, such as a deck you had just attached, was treated as a web address and failed. Names with spaces, and other everyday characters, now open normally.

• bug fix
  Retired Tools Left In The List

  When a tool is replaced, its entry could stay in the tool list after updating the app and could still be switched on, even though nothing was behind it any more. Replaced tools are now removed when the app updates, and a tool you had switched on is carried over to whatever replaced it instead of quietly disappearing from your selection.

• bug fix
  Long Data Series Came Back Empty

  Asking for a long run of data, such as five years of daily prices, could come back naming its source and containing nothing at all, with nothing in the reply to say the data had been dropped. This affected ordinary chat as much as scripts. As much of the series as fits is now kept, and the reply says how many entries it is showing out of the total.

• bug fix
  A Stalled Reply No Longer Hangs

  When a reply stopped arriving partway through, the app could sit waiting for eight minutes or more, because the time limit you can set was never applied to a reply as it comes in. It now applies everywhere, so a stalled reply gives up promptly and can be tried again. This covers chat, the mini apps and the Agent.

• bug fix
  A Batch Of Edits Lost To One Tag

  When the AI changed a file in several places at once, a single mistake in the way it marked up one of those changes had the whole batch refused and every edit in it lost. That mistake is now repaired and the edits are applied.

• bug fix
  A Copy Of A Built-In Model That Could Not Be Removed

  Private Cloud Compute is built into Privacy AI rather than something you set up, so it cannot be deleted by accident. Duplicating it, though, produced a second copy that was just as permanent and could not be edited or removed either. Built-in models can no longer be duplicated, and Refresh Servers now brings Private Cloud Compute back on a device that has just become eligible for it, without restarting the app.

Local Models

• improvement
  A Faster Local Engine

  The llama.cpp engine behind GGUF models has been brought up to version 0.4.1, taking in 310 upstream changes. Most of them land on Apple hardware: attention is now tuned per chip across the M1 to M5 and A18 families, a new sparse attention path speeds up long prompts, and the newest M5 and A19 chips use Apple's Metal 4 tensor units. Several memory leaks that grew over a long chat are fixed, so sessions stay steady. Reading images and documents gains a DeepSeek V4 Flash Vision encoder, and four new text architectures can now run on device, including DFM Mimir 1B and Tencent Hy 4.

• improvement
  A Smarter MLX Engine

  The MLX engine behind on-device MLX models has been updated to version 20260918. When a model writes a tool call in a style the app was not expecting, the request is now understood anyway, so the right tool still runs instead of the call turning up as plain text in the reply. A call whose values do not match what the tool accepts, or that was cut off partway through, is no longer run with pieces missing, and one written inside the model's own thinking notes, a quotation or a code block is treated as text and left alone. Emoji made of several parts, such as flags and accented letters, no longer repeat themselves while an answer is being typed out, and ordinary text sitting just before a tool call is no longer held back until the answer finishes. Loading a model no longer ties up the app's shared background capacity, generation runs on its own dedicated thread, and ParoQuant models load much faster after the first time.

Local Audio

• feature
  Live Speech Recognition In 33 Languages

  A new downloadable recogniser from NVIDIA writes words down as you speak instead of waiting for you to pause, so dictation keeps up with you in real time. It covers 33 languages and takes about 756 MB.

• improvement
  A Broader Speech Engine

  The mlx-audio engine behind on-device speech has been brought up to version 20260918, taking in 48 upstream changes and thirteen new families of speech model. Listening gains Whisper in every size, a model that writes down what was said and labels who said it in one pass, streaming recognition built for a live microphone, and four more recognisers including one made for low latency. Reading aloud gains five new voice systems, among them cloning a voice from a short clip, controllable pitch and speed, bilingual Chinese and English, and Japanese where you describe the voice in words instead of supplying a recording. Working out who spoke when is now about 1.8 times faster while listening, a loud or clipped recording no longer comes back with nothing written down at all, and Qwen3 recognition is more accurate.

• improvement
  Clearer Voices And Fuller Transcripts

  The FluidAudio engine behind Kokoro voices, Parakeet transcription and speaker separation has been brought up to version 20260918, taking in 216 upstream changes. Kokoro now runs its whole chain on the Neural Engine and kept all 54 English voices, and it reads English more carefully: a pronunciation dictionary is consulted before the neural speller, initialisms like NASA are read out as letters, possessives and hyphenated words are handled properly, and numbers and times are spoken the way you would say them. Two audio flaws are gone, a harshness in the high frequencies and an overlap error that made the joins between chunks audible, and choosing a different voice or speed now takes effect on the next sentence instead of waiting for the engine to be reloaded. Long recordings transcribe more completely, because a series of fixes to the way audio chunks are stitched back together stops words going missing at the joins and at the very end, and separating speakers is more accurate on recorded audio. An interrupted download now resumes from where it stopped instead of starting over, a stalled one is spotted in seconds rather than minutes, and cancelling no longer throws away a model that had already finished. Two speech recognition models were retired upstream and have been removed, multilingual Qwen3 and Mandarin, along with the older Kokoro voice chain that ran on the processor; Parakeet and Cohere Transcribe are unaffected.

Local Decisions

• feature
  A Model That Decides Instead Of Writing

  A new kind of on-device model answers a question about a piece of text and tells you how sure it is, rather than writing a reply. Ask whether an email is waiting on an answer, which of a dozen folders it belongs in, or how urgent it sounds, and the answer comes back as a probability for each choice, worked out in a single pass with no text generated at all. It handles three kinds of question: pick one from a list, yes or no, and a rating on a scale. Laya Multilingual comes in two reading lengths, about 190 or 570 words at a time, each around 485 MB, and both run on the Neural Engine, so a decision finishes even after you leave the app. Find them under Models and Services, Local Decisions, where you can download either and choose which one other features should use. For now they are reached through Privacy AI's own local API rather than from a chat, so this is groundwork: the app will start using them for jobs like sorting your inbox in a later release.

Agent

• feature
  Sub-Agents On Remote Models

  Handing a subtask to a sub-agent now works with remote models, as it already did with on-device ones. The sub-agent works in its own context and hands back only its answer, so a long piece of delegated work does not crowd the main run, and it can use a different model from the one running the agent: set a Lightweight model on the preset and every delegated subtask goes to it. Each subtask leaves its own report in the run's folder, and the run's stats list every sub-agent by name with its model, tokens, cost and time.

• bug fix
  A Failed Request No Longer Counts As A Turn

  When a reply failed to arrive, the run treated the turn as done and moved on, so a spell of network trouble could use up a run's turns with nothing completed. In one run seventeen requests failed and half of its working time went with them, unnoticed. A failed request is now recognised as one, so the run tries again and tells you.

• bug fix
  A Saved Result That Was Not Complete

  A large result saved into the workspace could hold a fraction of what came back while being described as the complete copy. A monthly report arrived with 34 of its 428 rows, and the run spent most of an hour fetching the same figures a day at a time, finishing with three quarters of them. Saved results are now complete, and that same job takes a small fraction of the steps.

Python

• feature
  Statistics And Machine Learning

  Python on your device now includes scipy and scikit-learn, on both iPhone and Mac. A script can run a hypothesis test, fit a curve, filter a signal, or train and score a model without installing anything, and the heavy arithmetic runs on Apple's own maths engine. Two things worth knowing: models train on a single core, so keep datasets modest, and the deprecated scipy.odr module is not included.

• feature
  Regression And Forecasting

  statsmodels is now built in, on both iPhone and Mac. A script can fit a regression and get p values and confidence intervals rather than just a line of best fit, write the model in the familiar formula style, test whether a series is stationary, and forecast one that changes over time with ARIMA. seaborn joins matplotlib for statistical charts.

• feature
  Financial Analysis And Backtesting

  Three finance libraries are now included. TA-Lib brings 161 market indicators, among them moving averages, RSI, MACD and Bollinger Bands, all running as compiled code rather than as a script. backtesting.py runs a trading idea over historical prices and reports what it would have returned, its worst loss along the way and every trade it took. quantstats turns a set of returns into the usual performance figures such as Sharpe and Sortino ratios and maximum drawdown. Prices are not fetched for you, so bring your own from a file, an API or the workspace.

• improvement
  Opening A File By Name

  A script could save to "report.docx" by name but could not open one, because a plain name was looked for somewhere you cannot reach. Opening now lands in the chat workspace just as saving does, for documents, slides, spreadsheets and PDFs, and the same holds for the charts and reports the new analysis libraries write.

General

• feature
  Python Can Read And Write Audio

  Python on your device could create sound from scratch but could not open an mp3, or produce a file your phone plays natively. A new audio module uses the app's own audio engine to read and write mp3, m4a, wav, aiff, caf and flac, with nothing extra to install. Reading gives you the samples and the sample rate; writing takes them back. mp3 is read only, because Apple ships a decoder and no encoder on any of its platforms, so save as m4a instead.

• feature
  Install Extra Python Packages

  The Install button in the skill tester now works. The bundled installer was unusable before, so any package that was not already included was simply out of reach, and the button could never succeed. Pure Python packages now install on demand. Packages containing compiled code still cannot be added on the device.

• feature
  More Python Libraries Included

  pandas joins numpy and matplotlib for tables and analysis. PyGithub, the Google API client, boto3, asyncssh, and Docker and GitLab clients let a script talk to those services directly. orjson, pydantic core, jiter and rpds make reading and checking JSON faster.

• improvement
  Chinese And Japanese Chart Labels

  Chart labels written in Chinese or Japanese came out as empty boxes, because no font covering those characters shipped with the plotting library. A font is now included and used by default, so the labels render without any extra code in your script. Korean is not covered by that font.

• improvement
  Relative Paths Land In Your Workspace

  A script that opened "notes.txt" or saved a chart to "chart.png" wrote it somewhere you could not reach. A plain relative path now resolves into the chat workspace, both for the script's own file handling and for pandas, matplotlib and Pillow.

• improvement
  The Same Python On iPhone And Mac

  Several libraries were present on Mac and missing on iPhone, so the same script could work in one place and fail in the other. Every bundled package is now present on both.

• improvement
  App Store Connect Tool

  A long list from the App Store Connect tool is now bounded, so a large account no longer floods a reply, and a list that reaches its limit no longer ends in a half written row. Win back offer pricing is available. The tool's settings screen is usable with VoiceOver and appears in your own language.

• bug fix
  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.

• bug fix
  A Cut Off Result Now Says It Was Cut

  When a tool produced more than fits in a reply, the result was shortened silently. The AI could then read half a table as though it were the whole thing and state a wrong conclusion with confidence. A shortened result now says so, in every mode. Five related fixes ship with it: a script that sleeps past the time limit now ends as a timeout instead of reporting success, and routine warnings from the numerical libraries no longer bury the output you asked for.

Local Models

• bug fix
  Qwen3.5 Models Load Again

  Every Qwen3.5 model stopped loading after the engine update in 3.1.0 and reported a model loading error instead of answering. This covered the 0.8B, 2B and 4B downloads and the community versions of them. The files recorded one piece of their vocabulary information in a form the updated engine had started to refuse. Those models have been published again in the form it expects, and the engine now accepts both, so a model already sitting on your device works again without downloading it a second time.

General

• improvement
  Document Outline Read Aloud

  The outline of a document now tells VoiceOver how deep each entry sits, so a sub-section is no longer read exactly like a chapter, and every entry says that it jumps to that part of the document. The two loading indicators in the reader are now named and report how far along they are instead of staying silent.

• bug fix
  Workspace Web Page Preview

  Tapping a web page in a chat's or an agent run's workspace files reported that the server could not be found instead of showing the page. Pages the AI creates, such as a slide deck or a report with its own images, now open properly, with their pictures loaded and the page's own layout and zoom left exactly as it intended.

• bug fix
  Text Above The First Heading

  A document whose first heading came partway through lost everything above it. A two page scanned form showed only its second page: the title, the whole table and both paragraphs of small print were missing from the screen even though the file itself was complete. Everything before the first heading is now shown, and it appears in the outline under a name taken from its own first line. This affected every place a document is read, including scans, chat attachments, workspace files, agent reports and saved web articles. Separately, a line starting with a # inside a block of code was mistaken for a heading, which cut the code block in half and added a stray entry to the outline. Code blocks now stay whole.

• bug fix
  Translated Search Boxes

  The search boxes for choosing a model, finding a server and picking from a list showed their prompt text in English even when the app was set to another language. They now follow your chosen language like the rest of the app.

Doc Scanner

• feature
  Doc Scanner

  A new app that turns a PDF, a photo, or a camera scan into clean text you can read, search, share, or chat with, and it all happens on your device. Every page is analysed twice. First a small layout model finds each element on the page and works out the order a person would read them in. Then each element is read on its own by a vision model, with a prompt matching what it is, so a table comes back as a table, a chart as a chart, and an equation as an equation. Reading a whole page in one go is what makes an AI invent text that is not there, so Privacy AI does not do it. Expect roughly a minute per page, because a dense page holds sixteen or more separate elements and each one is a separate read. A scan that is interrupted picks up exactly where it stopped, and nothing is read twice.

• improvement
  Doc Scanner Reads More Accurately

  Two problems were making scans worse than they had to be. The reading model was picking each word at random from its ten best guesses instead of taking its best one, which put stray symbols at the start of paragraphs and occasionally turned a whole paragraph into nonsense. And the page analysis was set stricter than it should have been, so faint text was thrown away before it was ever read: on one paper an entire address block and a journal masthead simply vanished. Both are fixed. An unreadable crop, such as a diagonal watermark, no longer fills a block with one repeated character, and a scan that is paused and resumed now produces exactly the same document as one that runs straight through. Measured on a real paper, the finished text went from 93 percent to 97 percent agreement with the document's own words. The first fix helps every on-device model, not just Doc Scanner's.

• improvement
  Three New Doc Scanner Settings

  New scans now offer a detail level, Standard or Thorough. Thorough is the default and is what finds faint straplines, addresses and captions; Standard keeps only what the page analysis is confident about, which is tidier on plain prose. Two more switches decide what goes into the finished document: whether to include the charts and images as picture files alongside the text, and whether to keep page numbers. Both can be changed after a scan has finished and rebuild the document from what has already been read, without scanning anything again.

• improvement
  Doc Scanner Page Titles

  A title printed in the same words on more than one page is now treated as a title on every page. When the page analysis was unsure and called one of them a figure caption, that page came out in italics with no heading at all, and a two page form ended up with no heading above its own first page. Privacy AI now compares a page's opening line with the titles found elsewhere in the same document and renders it the same way on both. Genuine figure, chart and table captions are unchanged.

Email Assistant

• improvement
  Search And Filter A Report

  A run report covering ten or more emails now has a search box and a filter above the list. Type part of a sender, a subject, a summary line, a category or a folder name and the decisions narrow as you type, and the filter picks out just the ones that need a reply, that were suggested, that went to junk, that are private, or that you undid, or a single category. A count under the box tells you how many of the run's emails you are looking at, and the search matches accented sender names and full width Chinese subjects however you type them.

• improvement
  Actions Within Reach

  What the assistant created for you, calendar events, reminders and notes, and what it learned from your mail now sit directly under the summary instead of below every decision. On a run covering dozens of threads you no longer scroll past all of them to undo something the assistant did, or to apply something it only suggested. VoiceOver reads the report in that same new order.

Local Models

• improvement
  Faster, Steadier On-Device Models

  The engine that runs models on your device has been updated with 391 changes from llama.cpp. Generation is faster on every Apple chip from M1 through M5, because the engine now carries settings tuned for each one rather than a single compromise, and this is the biggest speed change in the update. A memory leak that grew with every reply has been fixed, so a long conversation no longer creeps upward in memory. Images are now prepared exactly the way the models were trained to expect, which makes every on-device vision model read more accurately, Doc Scanner included. A video prompt costs about half what it did, because frames from the same clip are grouped back together instead of being treated as unrelated pictures. A crash when trimming a very long conversation with a compressed memory cache is fixed, and compressed caches now work correctly alongside fast attention. Prompts also start sooner, thanks to a fix to a chat template step that grew slower the longer the conversation was.

• improvement
  Eleven New Model Families

  Four of them are small enough to run on a phone: Granite SWA, Granite Switch 3B, Nanbeige 4.2 3B, and Pocket TTS. The engine also gained BailingMoE3, Kimi K3, MiniMax 01, Qwen3.8 Flash Next, Muse Glimmer, dots3-note, and speculative decoding for GLM-4.5-Air. New picture and sound understanding arrived with them: dots3-note reads both images and audio, Muse Glimmer reads images, and Pocket TTS can generate a voice.

• improvement
  Faster MLX Model Loading And Generation

  The engine that runs MLX models on your device has been updated with 54 changes. Models now load up to 1.8 times faster, because a model file is read in several pieces at once instead of one after another. Qwen3.5 and Qwen3 mixture of experts models generate text faster through fused internal steps and a shared expert selection step, and short prompts start answering sooner on Qwen models. Picture prompts on Qwen3 VL, Qwen3.5 VL and GLM-OCR are now processed in bounded pieces on every device, so a large image no longer exhausts memory, and a follow up question in a picture conversation no longer fails. Tool calls are more dependable too: a whole number such as a limit of 1 is no longer turned into true or false, a value containing brackets, commas or colons is no longer cut short, and Gemma tool calls with nested values are read correctly. Several Qwen3 VL downloads that previously failed to load, reporting that no weights were found, now load correctly.

• improvement
  New MLX Models And Rerankers

  MLX gained Muse Glimmer, a 30 billion parameter model that reads images and calls tools, along with Helium and TranslateGemma for translation. Document reranking models arrived as well, Jina Reranker v3 and BGE, which sort search results by how well they answer your question.

General

• feature
  Cache Details For Each Reply

  The details under a reply now show how much of the prompt the provider served from its cache, how much was added to the cache, and what that saved you compared with paying the full input price. Long chats and repeated questions on providers that cache show a large cached share, and a reply from a provider that reports nothing shows a dash rather than a misleading zero.

• improvement
  Health Data Tool Removed

  The get_health_data tool and the Apple Health integration behind it have been removed. Privacy AI no longer asks for or reads any Apple Health data, and the Health permission no longer appears in the app.

• improvement
  Costs Now Account For Cached Prompts

  Many AI providers charge much less for the part of a prompt they have already seen, and Privacy AI now uses that cheaper rate when the provider reports it and the price is known. Costs shown per message and in Statistics are noticeably lower for long chats and agent runs on those providers, and a spending limit now stops you at your real spend instead of well before it. When a provider reports cached tokens but no cached price is published, the old full-rate figure is kept so the number is never optimistic. Previously recorded costs are left as they were.

• improvement
  More Dependable Tool Use

  The descriptions and input rules the AI receives for every tool were audited and corrected, so tools are chosen more accurately and called with the right values the first time. Two tools that were being offered with no inputs at all now work as documented, and when a tool call fails the AI is told what actually went wrong so it can correct itself instead of giving up.

• improvement
  Clearer File Tools

  The single workspace tool the AI used for saving, listing and deleting files has been replaced by three plainly named ones: write a file, list files, delete a file. The AI no longer has to pick the right mode on a general purpose tool, which is where it most often went wrong and lost a turn correcting itself. Saving, listing and deleting all work exactly as before.

• improvement
  A Stronger Python Tool

  The built-in Python runtime now supports the latest version of packages: numpy, Pillow, and regex work on iPhone and iPad. A script that ends with exit() is now reported as a clean finish rather than "Unknown error".

• improvement
  Better At Reading Blocked Pages

  When a web page answers with a block or a verification wall, Privacy AI now recognizes it from the server's own response instead of guessing from the wording on the page. Articles that merely discuss captchas or sign-in walls are no longer thrown away, a page that was never really read is no longer listed as a source, and a very short page comes back with a note about its length instead of as a failure.

• improvement
  Simpler Delete Confirmation

  Confirming a permanent delete now asks you to type the word DELETE instead of retyping the item's exact name. Items named with emoji, or in a language your current keyboard cannot type, can finally be removed without switching keyboards, and lowercase works too. The message still names the item so you can see exactly what is about to go.

• improvement
  Python Scripts Can Use Their Own Files

  A Python script stored alongside others, such as the scripts inside a skill, can now import them the way Python normally does. Splitting a long script into a few files finally works. Editing one of those files and running the script again now picks up the change, instead of quietly using the previous version. Running the Python tool with both inline code and a file now runs the code and tells you the file was ignored, instead of refusing the call outright.

• improvement
  Charts From The Python Tool

  The built-in Python tool can now draw charts, on iPhone and iPad as well as Mac. Ask for a plot of data in your workspace and the chart is saved as an image you can open, share, and keep. Python also gained packages for spreadsheets, symbolic math, network graphs, and turning web pages into clean text, so more data questions can be answered on device without downloading anything.

• bug fix
  Sub-Agents On On-Device Models

  Handing a subtask to a sub-agent now works with on-device models. The sub-agent runs on a background copy of the model, and the previous one is freed first, so a local model can delegate without trying to hold two models in memory at once.

• bug fix
  Tool Steps Stay Visible After Rethink

  Asking for a rethink no longer removes the tool steps shown under the previous reply. They stay visible so you can see what was done, and that reply can be rethought again.

• bug fix
  Captions From Short Video Links

  Fixed subtitle downloads from some short video sites coming back empty because of a language mismatch, so a shared video link produces a transcript again. If the first read of the page fails, Privacy AI now keeps going and can still recover the transcript instead of abandoning the whole request.

• bug fix
  Flight Lookups Work Again

  The flight tool was offered to the AI with no description and no inputs, so it was rarely chosen and could not be called correctly when it was. Asking for a flight's status, delay, gate, terminal, or baggage carousel by flight number now works, as does asking for the arrivals or departures board at an airport.

• bug fix
  More Reliable Shell Commands

  The AI was working from outdated notes about the shell tool. It was told it could chain commands with && and || on iPhone and iPad, where those are not supported, and told the cut command was unavailable when it works fine. Shell commands now succeed more often on the first try, and these notes can be corrected in future without waiting for an app update.

• bug fix
  Chained Shell Commands On iPhone And iPad

  Chaining shell commands with &&, || and ; now works on iPhone and iPad, in the AI's shell tool and in the Shell view. Writing a command such as saving text to a file and then reading it back used to run only the first half and report success with nothing to show; now every step runs, a step that fails stops the ones that depend on it, and the output of the whole chain comes back. Commands that only mention && inside quotes are still passed through untouched. A blocked command can no longer slip through by hiding after && or a semicolon.

• bug fix
  Complete And Clean Shell Output

  Short shell results, such as the current folder or a line count, could come back empty and then reappear under a later command, and internal status messages could be mixed into what a command printed. A command that had worked perfectly could also come back with pages of internal tracing listed as errors, which made a success look like a failure and wasted part of the AI's reading budget on every command. Shell output on iPhone and iPad is now complete and contains only the command's own text, so the AI reads the right result the first time. Genuine warnings and error messages are still reported, whether or not the command succeeded. The same fix applies to the Shell screen.

• bug fix
  Tool Requests Rejected By Some Providers

  On providers that use the newer OpenAI response format, a chat with the automatic tool selector turned on could fail outright before the AI even began, because two tools described one of their inputs in a way that provider refuses. Those inputs are now described the way every provider accepts, so those chats answer normally again. The video editing and Python tools still accept exactly what they accepted before.

• bug fix
  No iCloud Access When Turned Off

  With iCloud Drive switched off, Privacy AI now leaves iCloud completely alone at startup. The Privacy AI folder no longer reappears in iCloud Drive after you delete it, and the iCloud Data Found question is no longer asked on its own. Checking for data left in iCloud by an earlier install is now something you start yourself, under Settings, Storage, Manage iCloud Leftovers, and cleaning it finds everything this time and removes the folder from Files for good.

Statistics

• feature
  Cache Statistics And Savings

  Statistics now shows Cached Input and Cache Saved in Today, This Month, and Overall, so you can see how much of your input was served from a provider cache and what that saved you. The shared statistics report carries the same figures. A period where no usage reported any cache activity shows a dash instead of zero, so nothing reported is never confused with nothing cached.

• improvement
  Cache Detail By Model And Provider

  Today's model table now has its own Cached and Cache Write columns, with the hit rate under each count and the amount saved beside the cost, so you can tell which models are actually benefiting from caching. Two new charts, Provider Cache and Model Cache, appear on all three tabs and show cached against uncached input in one bar, so a model that served one million of ten million from cache no longer looks the same as one that served one million of one million. The table also scrolls sideways at every window size, so the cost column stays reachable in a narrow window or a split view.

• improvement
  An About Page For Statistics

  A new About item in the Statistics menu explains what the numbers mean: why costs are estimates rather than a bill, what each period covers, what a dash in a cache column means, and that your records stay on your device and sync only through your own iCloud account.

• improvement
  Statistics Charts Read Aloud

  Every chart on the Statistics screen now describes itself to VoiceOver as one summary of its rows instead of staying silent, and each row of the model table is read as a single phrase naming every column it shows.

• bug fix
  Refresh In Statistics

  The Refresh item in the Statistics menu did nothing at all. It now reloads the figures for the tab you are viewing, which matters when a background agent, an extension, or another device records usage while the screen is open.

• bug fix
  Even Statistics Cards

  The Media Generations card stood taller than the cards beside it because its title wrapped onto a second line. Card titles now shrink to fit instead of wrapping, so every card in a row is the same height, in every language.

• bug fix
  Statistics Headings In Your Language

  The Provider Usage, Provider Cost, Model Usage, and Model Cost headings, and the messages shown when a chart has nothing to display, stayed in English whatever language the app was set to. They now follow your chosen language.

Agent

• feature
  Agent

  A new agent that works toward one goal over many turns instead of answering in a single reply. Give it a goal and it writes a plan, calls tools, keeps its own folder of files, and delivers a written report you can read, share, and reopen later. It suits work that has to be broken down first, such as deep research, a multi-leg trip, or a comparison of competing options.

• improvement
  Agent Plans Survive Longer Runs

  In the Agent mini app the plan is now read back from the run's own folder whenever it is not in memory, so a long run can keep updating the plan it wrote earlier instead of starting a new one and losing steps. Each run also carries its own reasoning forward between steps, gets a review of the report it actually finished rather than an early draft, and remembers a page that was blocked so it does not ask for it again and again. The list of sources at the end of a report now separates the ones the report cites from the ones that were only opened.

Skills

• feature
  Skills

  A skill is a folder of written instructions, plus any scripts and data it needs, that teaches the AI how you want a particular job done. Your skills live in one library under Settings, Tools and MCP, Skills, and each one becomes a tool the AI can call by name, so a library of thirty skills costs nothing until one is actually used. Skills work in ordinary chats and in the mini apps, and the AI picks up the full instructions only at the moment it calls one.

• feature
  Import Skills From Anywhere

  Add a skill by pasting a link to a public repository, by choosing a zip file or a folder, or by writing one yourself in the built in editor. A link that points at a single folder inside a repository now fetches just that folder instead of downloading the whole project, which is far quicker and works on large repositories. The screen shows exactly what was found, and where it came from, before anything is added to your library. Skills written for other AI coding tools work here unchanged.

• feature
  Files And Scripts Inside A Skill

  A skill can carry its own Python scripts, data files, and documents, and they are now managed in the app. Upload files, tap one to read or edit it, and delete the ones you no longer need. Text files open in a full screen editor and everything else opens in a preview, so a skill no longer has to be maintained outside the app.

• feature
  Let An Agent Write A Skill's Scripts

  Create Scripts hands your skill to an agent that writes the Python it needs, runs it, and offers the finished files back for you to review before anything is saved. It is the fastest way to turn a page of instructions into a skill that actually does the work.

• feature
  Test A Skill Before You Rely On It

  Two checks sit at the bottom of every skill. Check compatibility looks at the Python a skill imports and tells you which packages this device actually has, what to use instead of the ones it does not, and offers to install what it can. Test with AI runs a real conversation with only that skill switched on, so you can see whether a model chooses it and follows it. Either check can offer a reviewed one tap fix, always showing you the change first.

• feature
  Skills For Your Agents

  An agent preset can now be given skills from your library. Each one is copied into the preset so a later edit to the library cannot change a run that was working, and the preset tells you when its copy is out of date, when you have edited it locally, or when the original is gone. Updating from the library is one tap, and it asks first if you would lose your own edits.

• improvement
  Skills Written For Other Tools Still Work Here

  Skills shared online are usually written for a desktop computer, so their instructions often say to run a script from a command line, which iPhone and iPad cannot do. When a skill like that is used, Privacy AI now hands the AI the exact line to run instead, naming the real file, so an imported skill works on the first try rather than failing halfway through. The skill's own text is never rewritten behind your back.

Email Assistant

• feature
  Chat About One Email

  You can now start a chat about a single email thread from its report. Ask what a message really means, work out how to respond, or dig into the details, with the thread already in front of the assistant. Each chat keeps its own transcript and its own choice of tools, which start switched off so nothing runs unless you turn it on. The chat is stored with the report it came from and goes away when you delete that report, and a private thread is answered by your private model, the same one that drafts its replies.

• feature
  Alerts On Events And Reminders

  Calendar events and reminders created from your mail now come with an alert, so a meeting or a deadline no longer sits silently in your calendar until you happen to look. The assistant picks the lead time from the message itself, and an item with a date but no time alerts at 9 AM that morning. You can turn this off for an account under Actions in its settings, and nothing already in your calendar or reminders is changed either way.

• improvement
  Act On An Email While Reading It

  Opening an email from a report now puts the same actions menu in the navigation bar, in the app's usual style. Reply, chat, summarize, mark, change the category, undo a move, move to junk, or block the sender without going back to the list, and the details at the top of the email update right away to show what you did.

• improvement
  Write Your Own Reply

  The reply composer now opens with an editable draft, so you can type your reply yourself instead of generating one first and rewriting it. Save to Drafts, Open in Mail, and Copy all work on what you wrote, and asking for a generated draft when you have already typed something now checks with you before replacing it.

• improvement
  Sender Rules Stay Short

  Account settings now lists the three most recent sender rules, newest first, with a Show more button that reveals the rest. The assistant learns a new rule on most runs, so this list used to grow to dozens of entries and push Instructions and Test Connection far down the screen. Deleting a rule works exactly as before, and the next most recent rule takes its place.

• improvement
  A Tidier Report List

  An account now opens with only its newest month of reports expanded. Older months are collapsed behind a header showing how many reports each one holds, so an account with months of runs no longer scrolls on forever, and a collapsed month still shows a red dot when it holds a report you have not read. An older month can also be cleared in one go, confirmed by typing DELETE, and any report you pinned is kept.

• improvement
  Runs Resume Where They Stopped

  Processing all your inboxes no longer freezes when you leave the app or turn off the screen. iOS pauses the run after about thirty seconds, the work already done is kept, and the Live Activity tells you how many accounts were finished. Reopen Privacy AI and the rest carry on by themselves, with a Stop button if you would rather they did not, and no message is handled twice because mail the run never reached is still unread.

Local Audio

• feature
  Text To Audio

  Turn text or a document into an audio file using the speech models already on your device. Paste text or pick a PDF, ePub, or Word file, choose a voice, and Privacy AI reads it into a single WAV or M4A you can play, share, and keep. Nothing is uploaded, no subscription is needed, and a book-length conversion keeps its place if it is interrupted, so it carries on from where it stopped instead of starting over.

• bug fix
  Correct Name For File Transcription

  The Local Audio tab that transcribes an audio file was labeled File TTS, which suggested it turned text into speech. It does the opposite, so it is now called File ASR. Thanks to the reviewer who pointed this out.

• bug fix
  Pocket TTS Voices Play Again

  No Pocket TTS voice would speak in any language, whether you tried a preview or played chat text, even though the language pack showed as fully downloaded. The pack was being saved one folder deeper than the speech engine looks for it. Packs already on your device are corrected automatically the first time you open the app after updating, so nothing needs to be downloaded again, and new downloads are stored correctly.

• bug fix
  Kokoro Voices In More Languages

  Kokoro voices in Spanish, Italian, French, Hindi, Japanese, and Portuguese reported a model loading error when previewed and when chosen for speech. They now find the pronunciation files that come with the voice pack and speak normally. English and British voices were never affected. Thanks to the user who reported both of these in detail.

Local Models

• improvement
  Local Engine Update (llama.cpp b10333)

  The on-device inference engine was updated to the latest build, covering 234 changes since the last update. A GPU math fix corrects a subtle error that affected certain model shapes, so local replies are more accurate, and freeing a model that never ran on the GPU now releases its memory properly, leaving more room for the next one. Many newer models also load lighter because their draft acceleration layers are only loaded when they are actually used. This build adds MiniMax M3, Nanbeige 4.2, Laguna S-2.1, and GLM 5.2, along with new image and audio understanding from MiniMax M3 vision, GLM 5.2 Vision, MiMo v2.5 audio, and Parakeet speech. Multi page document reading runs faster, image resizing respects size limits more precisely, short audio clips no longer produce empty segments, and tool calling is sharper for the Qwen3, MiniMax M3, and DeepSeek V4 chat formats.

• improvement
  MLX Engine Update (20260810)

  The MLX on-device engine now reads long prompts about 9 percent faster on every MLX model, and Qwen3.5 and Qwen3.6 generate text up to 7.5 percent quicker. Image and video prompts on Qwen2.5 VL and Qwen2 VL are processed in bounded pieces, so large and multi image requests no longer exhaust memory, and OLMo 3 keeps its memory bounded in long conversations. Gemma models no longer drop the occasional character from a streamed reply, LFM2 tool calls read their arguments in full instead of cutting them short, and long Qwen3.5 and Qwen3.6 conversations hold better numerical accuracy. This update adds Qwen3 VL MoE, DeepSeek V2 and V2 Lite, the Hunyuan translation models, and Nanbeige 4.2, and lets Gemma 4 accept video directly.

• bug fix
  MLX Models Generate Again

  Several on-device MLX models stopped producing any reply after the last engine update, including Gemma 4, Granite 4 and LFM2.5. Privacy AI was asking the engine for a fixed memory window that those models do not support, so every attempt ended in an error before a single word was written. That request is gone, so each model now runs the way its own design intends and generates normally. As a side effect, long prompts keep their full instructions and tool descriptions instead of having the earliest part quietly discarded once a conversation grew past a couple of thousand words, so replies follow your setup and use tools more reliably in long chats.

• bug fix
  A Failed Retry Keeps Your Previous Reply

  If regenerating a reply failed, for example because there was not enough free memory to load the model, the reply turned into an empty red box with nothing to read. The previous reply is now kept, along with the arrows for moving between replies, and the reason the retry failed is shown as a note attached to it.

General

• bug fix
  Memory Profile Stays On Topic

  The memory profile that guides your chats now appears only when memory is turned on and a sector is selected, and it no longer blends details from unrelated sectors (such as work and family) into one paragraph. Each sector keeps its own separate profile, and upgrading clears any old mixed profile so a fresh, focused one is built as you use that sector.

Email Assistant

• feature
  Email Assistant

  A new mini app that triages your inbox for you. Connect any IMAP mail account with an app password, tap Process Inbox, and the AI reads your unread mail, tells you what matters, flags the messages that need your reply, and files the rest into tidy folders it manages under an AI label. You choose how much it does: summarize only, organize into folders, or also move obvious junk to your junk folder, and senders you have written to before are never junked. Every run produces a dated report where each decision can be undone with one tap, your corrections teach it your preferences, and what counts as important lives in a plain language instructions text you can read and edit yourself. Passwords stay in your device Keychain, message content is never stored, and nothing in your mailbox is ever deleted. Actionable mail can become calendar events, reminders, and notes, each listed in the report and undoable, and every report feeds a Knowledge Wiki for your account so the assistant builds a memory of your senders and projects over time.

Keyboard

• feature
  Name Your Pipeline Steps

  You can now give each step in an AI Keyboard pipeline its own name, so the step shows your label instead of a generic "Custom". This makes it easy to tell at a glance what each step does.

• feature
  Reusable Saved Steps

  Save any named step to a personal library and reuse it across your pipelines. Once saved, add it to any pipeline from the Add Step menu, and each one you add is an independent copy, so editing or deleting a saved step never changes the pipelines that already use it.

• bug fix
  Dismiss The Suggestion Bar

  The word suggestion bar can now be closed with an X while you type. It no longer sits on top of the pipeline picker, so you can switch pipelines at any time, and your next keystroke brings the suggestions right back.

Local Models

• feature
  Video on Local Vision Models

  MLX vision models can now understand video, not just images. Attach a short clip and the model reads it as a sequence of frames, and you can set how many frames are sampled, the frame rate, and the frame size per model to balance detail against memory.

• feature
  TurboQuant for Long Chats

  A new opt-in TurboQuant option lets an MLX model keep your entire conversation in on-device memory by compressing the cached values (keys stay high quality), so long chats use less memory with only a small quality trade-off. Turn it on per model in the model's Basic settings.

• feature
  Import MLX Models from ZIP

  You can now add your own MLX model from a ZIP file on your device, no download required. In the MLX section tap Import From Local, pick a zip that contains the model folder (config, weights, and tokenizer), and Privacy AI installs it and adds it to your model list, ready to chat fully offline.

• improvement
  Local Engine Update (llama.cpp b10091)

  The on-device inference engine was updated to the latest build for faster and more capable local models. Two-bit (Q2_0) models now run on the GPU on Apple Silicon instead of the processor, and newer chips get faster matmul for quicker replies. Image understanding is sharper too, with Qwen3-VL reading picture layouts more accurately. This build adds the Laguna code models and completes DeepSeek V4 support, brings smoother speculative decoding for faster answers, and strengthens stability and security when serving models.

• improvement
  MLX Engine Update

  The MLX on-device engine was updated with major memory savings for vision models: Qwen3.5 VL and Qwen3 VL now handle long and multi-image prompts using far less memory (avoiding out-of-memory crashes), and multi-turn image conversations stay consistent instead of drifting after the first reply. On-device generation now stops cleanly when the app moves to the background, preventing a crash during long prompt processing, and vision models no longer crash on the first tool call. Gemma 4 vision also handles multiple images in a single message correctly.

• improvement
  Engine-Aware Model Settings

  Local model settings now show only the options the model's engine actually uses, so MLX and GGUF models each display a tailored, honest set of controls, and every setting is tagged with the engine or engines that support it. MLX models also now respect the Top-K and Min-P sampling controls.

• bug fix
  No More Duplicate Imported Models

  Fixed an issue where a model imported from a local GGUF file could reappear as extra "Recovered from" copies in your model list after relaunching the app. Imports are now recorded once, and any duplicates created earlier are cleaned up automatically on the next launch, keeping your original model and its settings.

• bug fix
  Local Chat No Longer Fails When Memory Is On

  Local models with memory enabled no longer show a false error and now keep answering normally. When your chat model and memory model are both on-device, the app keeps your chat running and, if possible, uses Apple Intelligence to save memories in the background instead.

• bug fix
  Local Models Remember Your Conversation

  Small on-device models no longer claim they cannot remember what you said earlier in the same chat. A hidden bookkeeping line was being fed to the model as if it were an instruction, which primed it to deny having a memory. That line is gone, so local models now follow the conversation naturally and refer back to what you told them.

• bug fix
  Long Chat Setting Stays Put

  Changing how a local model handles conversations that outgrow its context window now sticks. Previously, if you adjusted this setting from inside a chat, it could quietly revert to the default the next time you opened the model. Your choice is now saved the moment you make it.

General

Keyboard

• feature
  Open Keyboard On Microphone

  You can now set the AI Keyboard to open straight into dictation instead of the key grid. Turn on "Open keyboard on microphone" in the keyboard's settings and the keyboard is ready to record every time it appears, so you can start speaking without hunting for the mic button. It stays off by default.

• feature
  Auto-Process After Dictation

  The keyboard can now polish your dictation automatically. Turn on "Auto-Process after dictation" and, the moment you stop recording, your selected pipeline runs on the transcribed text and hands back the finished result in place, with no Process tap. It stays off by default, and one-tap Undo restores your original words.

• feature
  Dictation Sound Cues

  The keyboard now plays subtle sounds as you dictate: one when recording starts, one when it stops and processing takes over, and one when your polished text is ready. This lets you follow the flow by ear without watching the screen, which is especially helpful with VoiceOver. Turn the cues off any time under Feedback in the keyboard's settings.

• feature
  Live Dictation Preview

  Watch your words appear as you speak. When you dictate with the AI Keyboard, your transcription now previews live in the recording panel so you can see it take shape, and the finished text is inserted the moment you stop. It turns on automatically when your chosen speech recognition supports live results, and you can turn Live Transcription off any time under Dictation in the keyboard's settings.

• improvement
  One Tap To Polish After Dictation

  After you finish dictating, VoiceOver focus now moves straight to the Process button and reads it aloud, so polishing your text is one clear tap with nothing to hunt for.

• bug fix
  Keeps Your Audio In High Quality While Dictating

  Recording with the keyboard no longer drops your Bluetooth headphones, AirPods, or the built-in speaker to muffled phone-call quality. Your music and VoiceOver stay at full quality while you dictate, and dictation now captures from the phone's microphone.

• feature
  Multi-Language Keyboard

  The AI Keyboard now types in your language. Tap the globe key to switch layouts across all 11 app languages, including English, Simplified and Traditional Chinese, Japanese, Korean, French, German, Italian, Spanish, Russian, and Brazilian Portuguese. Chinese Pinyin and Japanese offer full candidate input with a suggestion bar, Korean composes Hangul as you type, and the European layouts add accented characters (hold a key) plus a numbers and symbols page. The Chinese and Japanese dictionaries download on demand the first time you enable them.

• feature
  Inline AI Instructions

  You can now ask the AI to fill in part of your text as you write. Tap the new AI button on the keyboard, type an instruction such as "search today's SPX price" or "write a one line intro to Beijing", then tap AI again to close it. The instruction is replaced in place with a short, direct result the moment you close it, so the rest of your sentence stays exactly as you wrote it. Add as many instructions as you like in the same message, and there is no need to select any text first.

• feature
  Memory For The AI Keyboard

  The AI Keyboard can now use your saved memories. In the keyboard's settings, attach a memory sector so Process and inline AI instructions are grounded in your context, such as your identity, preferences, and writing style. Turn on Search Memories to also pull in specific saved facts relevant to what you are typing. It stays off until you choose a sector, so nothing changes unless you opt in, and your memories are read on your device by the app, never by the keyboard itself.

• improvement
  Keyboard Steps Keep Your Language

  Built-in keyboard pipeline steps such as Rewrite, Casual, Concise, and Summarize now stay in the language of the text you are transforming instead of switching to English. A pipeline that translates to Japanese and then makes the result casual now stays in Japanese. Your own custom steps still control their own output language.

• improvement
  Rewrite Your Full Selection

  Process now rewrites the exact text you select. Highlight the words you want to change, or use Select All for the whole field, then tap Process and the entire selection is rewritten in place. This reliably handles longer text, which could previously be processed only in part because the keyboard could see just a small portion of your field near the cursor.

Local Models

• improvement
  Local Engine Update (llama.cpp b9993)

  The on-device inference engine was updated to the latest build for faster and steadier local models. This adds support for many new model families, including Hunyuan v3, DeepSeek V4, MiniCPM 5, and the LFM2.5 embedding and retrieval models, plus Granite Speech Plus audio and Unlimited-OCR document understanding. Speculative decoding brings quicker replies on more models through MiniMax M2 and DFlash, alongside faster image and audio processing on Apple Silicon. An updated core improves performance and stability across the board and strengthens security when reading model files.

• improvement
  MLX Engine Update (mlx-swift-lm)

  The Apple MLX inference engine was refreshed for faster prompts, more accurate vision, and new local models. Gemma 3 now reads your prompt up to 2.6 times faster on Apple Silicon with identical results. Vision models are more reliable too: Qwen3 VL now reads dark and low-contrast images correctly, Qwen3.5 VL and Qwen2.5 VL produce cleaner answers on detailed images, Qwen2-VL places image regions with true spatial awareness, and more Gemma 4 vision models load without errors. This update also adds the Mixtral (mixture-of-experts) and Mamba2 language models and the LFM2.5 text embedders, lets generation stop on custom words, and adds a seed option for repeatable results. Gemma tool calls also pass their inputs with the right types for more dependable results.

Knowledge Wiki

• feature
  Knowledge Wiki

  Turn your documents into a personal, cross-linked knowledge base that the AI builds and maintains for you. Add your files and Privacy AI reads each one, writes clean summary pages, and links related people, topics, and ideas together, all kept on your device as plain Markdown you can also open in Obsidian. Create as many separate knowledge bases as you like, one for each project or subject.

• feature
  Add Sources From Anywhere

  Build a knowledge base from almost anything: PDF, Word, and EPUB documents, web pages, Markdown, text, and CSV files, plus photos, videos, and links you paste. Import from Files, Photos, the camera, or a web address, and the AI reads the whole document, not just the first few pages, so nothing important is missed. Add several sources at once and watch each one's progress as it is processed.

• feature
  Ask Your Knowledge Base

  Ask a question in plain language and get a clear answer drawn from your own sources, with every point linked back to the exact wiki pages it came from so you can check the details. Choose Quick for a fast, direct reply, or Thorough when you want the AI to weigh more of your material, and save any good answer back into your knowledge base for later.

• feature
  Interactive Knowledge Graph

  See how everything in your knowledge base connects in a live, interactive map. Related pages gather into color-coded topic clusters, stronger connections stand out, and you can pinch to zoom, drag things around, and tap any point to open its page. Zoom in to reveal more labels and explore the finer detail.

• feature
  Automatic Overview and Insights

  Each knowledge base keeps a living Overview page that the AI rewrites as you add material, so you get an up-to-date summary the moment you open it. Tap Graph Insights for an AI-written tour of the biggest themes and how your topics relate, with links straight to the pages.

• feature
  Wikis In Your Language

  Create a knowledge base in the language you want. Pick a language when you create it and the AI writes every page and answers every question in that language, across all supported app languages, while keeping your links and page names intact.

• feature
  Git Sync For Your Wiki

  Back up and sync any knowledge base with GitHub or another Git service. Publish a knowledge base to a private or public repository, pull updates on another device, or clone an existing repository straight into a new knowledge base. You can browse the full history of changes, compare versions line by line, and restore any earlier version, all from inside the app.

• feature
  Wiki Health Check

  Keep your knowledge base tidy with a one-tap Health Check. It finds pages that link nowhere, broken links, and missing pages so you can spot gaps, and an optional deeper review uses AI to flag possible contradictions and unverified claims.