Use your own model files If you already have a .gguf model — on a drive, in iCloud, or sent to you by someone else — you can now import it straight from the Files app. No download and no network, which matters if you're setting up a phone that has no connection at all. Imported models work exactly like downloaded ones: they show up in My Models, run entirely on-device, and can be deleted to free the space back up. Smaller iCloud backups Downloaded models are no longer copied into your iCloud backup. If your backups had quietly grown by several gigabytes, this was why — and this update fixes it for models you've already downloaded, not just new ones.

Search your chats The chat list now has a search bar. It looks inside your conversations, not just their titles, so you can find a chat by something that was said in it. Results show the matching text, so you can tell which one you meant without opening each.

Document Library (Pro) Save documents once and ask questions across all of them from any chat. Turn on "Search my library" and the relevant passages are found and given to the model, rather than the document being cut short. A long PDF used to be trimmed to a fraction of its length, with the model answering as though it had read the rest. Length is no longer the limit. Answers that used your documents show which file each excerpt came from, the passage itself, and how closely it matched — so you can check it pulled from the right place. You can also narrow a chat to specific documents when you already know which one holds the answer. Documents are read, indexed and stored on your device. Nothing is uploaded. Fixed Using one of your own photos as a starting point for image generation failed with an encoder error. Photos are now prepared at the size the model expects, and a portrait photo no longer comes out rotated. Also The tokens-per-second readout under each reply can be turned off in Profile. It was only ever measured on your device and never recorded or sent anywhere.

Fixed a bug where a long conversation could stop responding entirely. Once it hit the context limit, every following message failed the same way with no way to recover short of starting a new chat. Older messages are now set aside automatically so the conversation keeps going. Models also reserve memory for context more efficiently, which leaves more room on smaller devices, and attached documents are now sized to the model you have loaded instead of a fixed limit. Added Report a Bug in Profile — it drafts an email with your app version and device already filled in.

LaTeX mode writes maths, not documents Asking for a formula in LaTeX mode used to return an entire compilable document — preamble, packages and all — which arrived as raw source instead of a rendered equation. It now writes equations inline with the explanation, the way you'd expect, and keeps explaining in plain language around them. Honest web search DuckDuckGo intermittently blocks automated searches. Until now that looked exactly like finding nothing, so the answer quietly came from the model's own knowledge while appearing to have been freshly searched. The app now tells you when a search didn't run, and the model says so too, rather than presenting a confident answer as though it were current. Sources stay with their answer Search citations now sit directly beneath the reply they produced, and survive reopening the conversation. Previously they always showed the most recent search and were replaced the moment you asked something else — scrolling back to an earlier answer showed sources belonging to a later question, or none at all. Smaller download Removed 1.4 MB of unused resources left over from an older maths renderer.

Know what will run before you download it Every model now shows whether it runs comfortably on your iPhone, runs but uses most of your memory, or won't fit at all — measured against your device's actual RAM. The same badge appears in Discover, so you can tell at a glance instead of spending four gigabytes finding out. Newer models The on-device inference engine is roughly 1,900 builds newer. Qwen3.5 and other recent architectures now load and run properly instead of failing partway through. Reasoning models, properly handled Models that think before answering now collapse that thinking instead of showing it as the answer. Their reasoning no longer fiher, so long conversations stay coherent. And the response-length setting now limits the answer rather than the thinking — previously and its whole budget deliberating and never reach a reply. Formatting that holds together Headings, lists and bold text render as intended rather than leaking raw ** markers into replies, and line breaks stay whekdown now always renders, so the Markdown chat mode has been removed — it was a switch that appeared to do nothing. Fenced code appears as. Fixes • Em-dashes, accented characters and emoji courd • Headlines in Discover showed ’ instead of an apostrophe • Deleting a model while it was loading could • Models that failed to load did so silently — they now tell you why • A model could appear in your library with no • Some models were listed twice with conflicting memory requirements

Two things in this update, both from your feedback: Fixed a bug where parts of the onboarding could be cut off — on some iPhones, especially with Display Zoom enabled, content and the Next button could end up off-screen. Those screens now scroll, so nothing stays hidden. Lowered the price of the Lifetime unlock. A few of you said it was steep, and since MyAI runs entirely on your device, a cheaper one-time unlock made sense. Thanks to everyone who wrote in.

Feature Requests — there's now a link in your Profile to vote on what comes next in MyAI. Head to the Feature Requests page and let us know what you'd like to see built. Also added a Write a Review link in Profile if you've been meaning to leave one.

Fixed a crash when loading an image generation model while a chat model was active — the app now properly unloads before switching. Image model loading now shows a heads-up when the final stage slows down, so you know it's still working.

Document Chat — Attach a PDF, Word doc, Excel spreadsheet, or text file and ask your AI anything about it. Extracted and processed entirely on-device. Email Reply Mode — A new chat mode that drafts ready-to-send replies. Paste the email you received, tap "Open in Mail" to send it from your email app. Bug fixes and performance improvements.

v1.9 — Personas + Web Search AI Personas: Give your AI a personality. Tap the masks icon in the input bar to create a custom persona with a name, icon, and system prompt. Presets included, or write your own from scratch. Web Search: Research Mode searches DuckDuckGo before every response so any model can answer questions about current events and news. Off by default — your queries never include conversation content. Also in this update: — Input bar reorganized so the text field always has full width — Image generation models can now be deleted from My Models (swipe left) — Loading progress bar when initializing an image generation model

Chat Modes — Tap sparkle icon in the input bar to change how the AI responds. Seven modes: Markdown, LaTeX, Code, Concise, Step-by-step, ELI5, and Formal. Mode is saved per conversation. LaTeX rendering — Switch to LaTeX mode for math questions. Equations render as formatted expressions instead of raw symbols. Copy code — Every code block now has a copy button. Tap to copy the full snippet to your clipboard. File export — Tap the file card below any response to export it. Markdown saves as .md, code as .py/.js/etc., math as .tex. Share directly to Files, Notes, or any app.

Image Generation — generate images entirely on your device using Stable Diffusion. No internet required after the model downloads. Type a prompt, or upload a photo and reimagine it with adjustable strength. Fully private, zero cloud. Pro — unlock Arena mode, Voice tab, HuggingFace live search, Vision models, Image Generation, and unlimited saved models with a single subscription. Early supporters who installed before this update get Pro free, forever — no action needed. Qwen3 VL 2.5B — Alibaba's newest vision model is now available. Best-in-class at OCR, charts, and document reading, all on-device. Bug fixes and performance improvements.

Voice Mode — A dedicated Voice tab lets you talk to your AI and hear it respond, entirely on-device. Choose from any voice installed on your iPhone, including enhanced neural voices. Model Picker — Tap the model name to open a rich model sheet showing capability tags, quantization, size, and description for every downloaded model. Vision in Voice — Vision models now work in the Voice tab. Attach a photo and ask anything about it. Onboarding — New users now get a walkthrough of Voice, Arena, and Vision before their first chat. Re-engagement notifications — Optional nudges if you haven't chatted in a few days. Bug fixes — Fixed a crash when switching between vision and standard models. Fixed image attachments not displaying in chat after sending.

• Fixed an issue where the chat would scroll to a wrong position or show a blank screen when sending a message from the middle of your chat history • Fixed auto-scroll during AI responses — scrolling up to read previous messages no longer gets interrupted • Fixed a bug where your loaded model would unexpectedly unload after a few conversations, requiring you to reload it manually • Minimum iOS version is now 18.0

- Fixed a bug where selecting a model in the My Models tab caused chat to show an error when sending messages — model selection now happens directly in the Chat tab - Fixed a bug where scrolling up in a long conversation during generation would leave the screen stuck — you can now scroll freely while the model responds

Introducing Arena — set two AI models against each other and watch them debate, reason, and respond to one another entirely on your device. Pick from 16 personas (The Skeptic, Socrates, Tech Architect, CTO, and more), choose a topic or browse examples, set the number of rounds, and let them run. Sessions are saved so you can revisit them anytime. Choose your quantization — models now show every available variant (Q2_K through Q8_0) so you can pick the right balance of speed, size, and quality. Set a maximum output length in Settings to control how long responses can be. A new live stats card shows tokens/sec, token count, time to first token, and memory usage while the model is running.

Discover tab — browse the latest AI news from The Verge and TechCrunch, updated every 2 hours. Model categories — explore curated picks for Reasoning, Coding, General, and Creative use cases, all ready to download directly in-app. Trending — see what the AI community is downloading on HuggingFace right now. Learn tab — new educational content explaining how LLMs work, what quantization means, and how to pick the right model for your device.

Bug Fixes & Improvements - Fixed an issue where models could produce garbled or nonsensical output - Fixed a crash that could occur when switching between models - Added Gemma 3 support and improved compatibility with newer open-source models - Improved memory management for more stable inference on all devices