MyLLM 5.9.8 FIXES • Own server (Ollama): chats no longer lose track of earlier messages. A model MyLLM didn't know by name was treated as having a small memory, so earlier messages were left out; it now uses the context size you set. • Saying "proceed" or "yes" to a question the assistant asked works again: your previous message and its reply are always sent. • The assistant can now try up to three rounds of tools before answering, so "let me try something else" is followed through.

MyLLM 5.9.7 NEW • Edit your apps with the model on this iPhone: Shell now shows your saved apps as files in its apps/ folder. Ask MyLLM there to change one and it reads only the lines it needs, shows you each edit before it's made, and checks the app for errors afterwards. Works with no signal. • Only Shell can open your saved apps, and each change keeps the version it replaced, so undo puts it back. FIXES • Edit in Chat now asks first when your chats use the model on this iPhone, and offers to open the app in Shell or switch to your own server or a cloud model. Before, the phone model was asked anyway and could only say it didn't have the tools. • A request from Edit in Chat or Remix is no longer sent until you've typed the change you want.

MyLLM 5.9.6 FIXES • Long tasks on your own server (like editing a saved app) no longer get cut off part-way. MyLLM now makes sure every step leaves room for the model's reply. • If your server adds its own text to each request (a proxy adding search results, say), MyLLM now reads how big the request really was and leaves room for that too. • If a step is still cut off, it isn't run, and the assistant tells you that change wasn't made. • If a conversation is too full to go on, the assistant stops and says so. Start a new chat to carry on. • Big changes to saved apps are made in smaller steps, so they fit.

MyLLM 5.9.5 NEW: Phone helper (Settings → Privacy) • Personal data stays on this phone: when you chat with a cloud model or your own server, your files, email, contacts, calendar, photos and other personal data are read by the model on your iPhone, and only its answer is sent. • Condense long pages on this phone: long web pages are summarised on your iPhone first, so the cloud model reads less and costs less. Both are off until you turn them on. FIXES • Browse Hugging Face now checks each build against the memory iOS actually gives MyLLM on your iPhone, so it no longer suggests a model that downloads but can't load. • If a model is too big to load, the message now says how much memory it needs and how much is available, and suggests a smaller build. • The assistant is no longer told about tools it doesn't have in that turn.

MyLLM 5.9.4 Fixes for models added from Hugging Face. • Tapping Add vision on a model you added could also remove that model from your iPhone, so it disappeared from the chat's model menu. Each button now does only its own job, and Remove from List asks before deleting anything. • A model that finishes downloading now appears in the chat's model menu straight away. • A vision file left behind by a removed model is now cleaned up, freeing the space. If a model you added went missing, browse to it again in Settings → On-device models → Browse Hugging Face and download it.

MyLLM 5.9.3 Your MyLLMos apps on the Home Screen. • New widgets: MyLLMos App (one app) and MyLLMos Apps (four). One tap opens the app — and if you left it running, it carries on where you were. • In MyLLMos, touch and hold any app and choose Add to Home Screen… for step-by-step help, including a Shortcut with your own icon.

MyLLM 5.9.2 Apps that keep running, and report back. • Leaving a MyLLMos app no longer closes it. It stays running, paused, so you can check another app and come straight back — Shell keeps its whole session. A Running strip in MyLLMos lets you switch back or close them; up to four stay open. • Shell and Bin now come with MyLLM: the command line for your shared files, and the trash that lets you restore what you deleted. • When the assistant runs an action in one of your apps, like adding a flashcard, it now hears what the app did, the app's reply or the data it saved, instead of guessing whether it worked. • App makers: return a short summary from your myllmIntent.receive handler, or call intent.reply(summary), and the assistant is told. • The assistant can tell when its set of tools changes between messages, and isn't sent looking for tools it doesn't have. • The assistant explores your shared files with the same commands as the Shell app: pipes, patterns, version diffs and JSON queries. It still only reads; anything it wants to change goes through a tool that asks you first.

MyLLM 5.9.1 Tools you can count on. • A connected MCP server with lots of tools no longer crowds out MyLLM's own: its most-used tools come along, the rest are found when needed. Shared files and tool discovery are always there when the assistant is told about them, and it won't claim a save or search that didn't happen. • Files: text files, like a bot's notebook, can now be edited right in the preview. • Saves show as a receipt under the reply, like "Saved · Bots/Jarvis.md", and a failed tool says so. • Ask the assistant about MyLLM and it describes this version, and says where your messages are going right now: this phone, your server or a cloud provider. • MCP servers: a tool that changes something on the server now asks for your OK before it first runs in a chat (Auto-approve was shown but not applied before). Write tools are marked, and bots' scheduled duties can't run them. • Command Center: settings files are no longer listed as systems, and out-of-date readings are marked as not current. • The assistant is only told how to use the tools it actually has in that message, so it no longer promises searches, code or pictures it can't deliver. • The assistant can check exactly which tools it has in a message, and finds tools from everyday words like "reminders" or "translation". • Settings → Tool Permissions now really applies: deleting files, sending email, adding events or reminders, running Shortcuts and opening links ask once per chat before they run, and Deny means never (Settings → Built-in Tools → Tool Permissions). • Each permission card asks about one tool, says why it's asking, and what you tap is noted in the chat. The assistant sees the answer straight away: allow a tool mid-reply and it carries on, decline it and it stops trying. • Reminders: the assistant can check your reminders when it adds one, and won't add the same reminder twice.

MyLLM 5.9 Scan a room with LiDAR, and apps the assistant tests as it builds. • Scan a room: on iPhones with LiDAR (Pro models from the 12 Pro), MyLLMos apps can now scan a room into a floor plan, with walls, doors, windows and furniture measured to a few centimetres. • The new Room Scanner app turns a scan into floor area, wall lengths and ceiling height, with paint, flooring and skirting estimates. No LiDAR? It also works from measurements you type in. • Ask for a change to one of your saved apps and the assistant edits it, opens it out of sight, reads any errors, and fixes them before telling you it works. • It can also try a feature itself, such as tapping a button and checking that the list grew. • Small slips in copying the code, like the wrong indentation, no longer make an edit fail, and big apps are read a section at a time instead of losing the end of the file. • Working on an app now gets up to 30 steps in a normal chat, not only in Agent mode. • Switch models from the chat: tap the model name at the top to move between a model on this phone and any server or cloud provider you've already set up, without going into Settings. Each connection remembers its own model. • Fixed: testing an Anthropic key under Cloud (your key) failed with status 400 even when the key was right. • Seven more providers under Cloud (your key): DeepSeek, Z.ai GLM, Moonshot Kimi, Alibaba Qwen, Mistral, xAI Grok and Groq. • Fixed: newer Claude models answered every message with an error about temperature. When a model refuses a setting, MyLLM now leaves it out and asks again.

MyLLM 5.8.3 Phone models that use tools, and bots that react. • MiniCPM5 2B: quick and light, and it uses tools, so iPhones with 4GB now get a model on the phone that can check the weather, search the web and add reminders. • Gemma 4 E2B: the smaller Gemma 4 for iPhones with 6GB. It uses tools, and can read photos once its vision file is added. • Both were tested on the app's real tools before being added: they ask for the right tool, leave everyday questions alone, and answer from the result. • Duties that react: a bot's duty can now run when files change in a folder, or when one of your systems needs attention, not just at a set time. • Apps can hand things to bots: a MyLLMos app can send something to one of your bots, like today's tasks to the Planner. You always see it and tap Send first.

MyLLM 5.8.2 Memory that stays in its lane, and bots that stay in the chat. • Your assistant is back to engaging with the world: general questions get a fresh answer and a web search again, instead of drifting back to things it remembers about you. • Memories now come into play only when your message is about you, and at most a few at a time. • A handful of frequently used memories no longer ride along on every message. • When a memory is relevant, it's treated as background, never as a reason to skip looking something up. • @Planner (or any bot's name) now hands the chat to that bot: keep talking and it keeps answering, with no need to type @Planner again. • @myllm is the default assistant. Use it, or tap "Back to @myllm" above the message box, to go back. • Type @ to see your bots, with @myllm marked as the default. • The chat's title, the message box and a "Talking to" line show which bot is answering.

MyLLM 5.8.1 A calmer, clearer phone model. • New in Settings → On this iPhone → What the model reads: switches for Short memory, Memories in messages and App knowledge notes. • Short memory sends the phone's model only the recent part of a chat, so replies start sooner and the phone stays cooler. • Memories and app notes are no longer added to your messages to the phone's model unless you switch them on. • Learning from your chats now waits for ten quiet minutes or the charger, so your next message no longer has to read the whole conversation again. • When a message does have to read the whole chat again, the status line says why.

Your assistant now knows MyLLM. • Ask it about the app. "What's new?", "How do I turn on Attune?", "Can it work with no signal?": the model answers from MyLLM's own guide for the version you're running, on your phone's model or your server's, instead of guessing. • Meet your bots. If you missed it in 5.7: Personas are now Bots, each with a job, its own way of talking, a notebook and daily duties like a morning briefing. Start a message with @Planner to ask one directly. • A fresh App Gallery. Categories are compact shelves you swipe through with full app names, 'See all' opens a whole category, chips at the top take you to a category, your installed apps or updates, and every app has a proper page with its size, version, what it needs and more like it. • Pinned apps as a grid. Apps you pin in MyLLMos now sit on the chat's start screen in a grid, all in view, instead of a row you had to scroll. • Edit your bots. Every bot, ready-made ones included, has an Edit button on its page for its icon, colour, name, job and how it talks. • Steadier on-device chats. Bots no longer limit the phone model's tools, and the faster first message after opening the app now switches on only after Settings → On this iPhone → Self-test has checked it on your iPhone. • What's New shows what's actually new, and will from now on. Everything on your device stays on your device.

Meet your bots. • Personas are now Bots. Each one has a job, its own way of talking, and a notebook it keeps for itself. • Duties. Give a bot something to do every day, like the Daily Planner's morning briefing, and it's ready when you wake up. With the model on your phone it works quietly while the app is idle or overnight on the charger, then tells you what it found. • While you were away. A new chat opens with what your bots did since you last looked. • @mention a bot. Start a message with @Planner or @Research and that bot answers, then the chat carries on as before. • Bots now keep to their role for the whole chat. Before, a persona could slip back to the plain assistant after your first message. The model on your phone, sharper all round. • Ask for an app, get an app. Ask the model on your phone for a tracker, a quiz or a game and it builds one you can open and keep. While it writes, you see "Building your app…" instead of a wall of code, and apps now remember what you put in them after you close them. • Answers the question you just asked. If you stopped a reply, or it failed, the model on your phone could answer your earlier question instead of your new one, and read the whole chat again first. Fixed. • No more going round in circles. If a reply starts repeating itself, it is stopped early and written again, instead of running on for minutes. • Accents, hanzi and emoji come out right in on-device replies. Before, a character could occasionally turn into "?". • A faster first message. The model now keeps its reading of its own instructions between launches, so the first reply after opening the app starts sooner and the phone stays cooler. • Downloads keep going. A model keeps downloading when you leave MyLLM or lock your phone, and you're told when it's ready. Models are also left out of your iCloud backup. • Kinder to memory. When iOS runs short, MyLLM lets go of the least-needed parts first. If the app has closed twice in a row mid-answer, it now says so and offers a smaller model. • On this iPhone. Everything about your on-device model is now on one page in Settings: models, Attune, the privacy ledger, Ask your photos, quiet-time learning and the off-grid kit. There's also a self-test that checks your model in about a minute. • A clearer model list: the best model for your iPhone first, then models that use tools and read photos, then small chat-only ones. Everything on your device stays on your device.

Your phone's model can now lean towards how you like to be answered. • Attune (experimental). Describe how you'd like to be answered, in your own words, and the model on your phone leans that way: shorter, warmer, more formal, whatever suits you. No training and nothing added to your messages; the model simply reads, on your phone, and learns the direction. Settings → On-Device Models → Attune. • More like this, less like this. With Attune on, two small thumbs sit under each on-device reply. Rate a few and your next attune takes them into account. • Careful by design. Attune shapes manner, not facts: what it knows about you stays in Memory. It holds itself back whenever tools are in play so the weather, calendar and web search keep working, it never goes past a fixed limit, and if a reply ever starts going round in circles it is stopped and written again. • Gemma 4 can hand you an app. When Gemma 4 built a mini-app, MyLLM could lose it and show a wall of code instead. Fixed: ask for a tracker, a quiz or a lesson and you get a card you can open and save. Everything on your device stays on your device.

The model on your phone no longer leaves you guessing. • No more silent waits. When a phone model is reading a long conversation you now see how far it has got, and when a reply turns into a tool call it says so. These waits could look like a frozen app; thank you to the user who reported it. • Tools answer faster. After using a tool, a phone model used to re-read the whole chat before replying. It now carries straight on. Tool results are also sized to fit the phone's memory. • Stop always stops. The stop button now works while a tool is running too, even one waiting on your location or a slow website. • Deep Think really thinks. On Qwen 3.5 and Granite 4.2, Deep Think now switches on the model's own reasoning for that message, with a limit so it can't wander off for minutes. • Know before you install. Gallery apps that read photos or make pictures are badged, and the app page tells you if your current model can't do that, with how to fix it. • Internet Radio can show who's playing on most stations, and its dock now fills the whole screen. Everything on your device stays on your device.

The model on your phone can now act, see, and prove it kept your data to itself. • Tools on your phone. Qwen 3.5, Gemma 4, Hermes 3 and the new Granite 4.2 can check the weather, search the web, read a page, look at your calendar, set reminders and work with shared files. Look for the TOOLS badge in Settings → On-Device Models. • Photos, read on your phone. Add the vision file to Qwen 3.5 or Gemma 4 and ask about a photo without it leaving your device. • Ask your photos. Turn it on and MyLLM reads your photos and screenshots on your phone, so you can find "the receipt from the garage" or "the screenshot with the wifi password". Nothing is uploaded. • Private, and you can check. On-device replies carry a shield, and Settings → Privacy → On-device ledger shows every reply and any tool that went online for it. • Summarise on This Phone. Share a page, text or a picture from any app and a model on your device summarises it, whatever your chat is set to use. Also in Siri and Shortcuts. • Off-grid. Lose your signal and MyLLM offers to carry on with the model on your phone. The new Off-grid kit shows what works with no connection. • Learns in quiet moments. On-device chats now build your memory too, read back when the phone is idle or charging. • Better mini-apps. Apps can ask for an answer in an exact shape and get it, even from the smallest model. • Two new models: Granite 4.2 3B and Qwen 3.5 9B (for 12 GB devices). Everything on your device stays on your device.

Better answers from the models you already have. • Qwen 3.5 answers straight away. It was quietly "thinking" before every reply, which its makers switch off by default. Replies now arrive in a fraction of the time, with the same quality. • The models you get when you first install (LFM2 and Qwen 2.5) now use the chat format they were trained on. Apps that ask them for a recipe, a meal plan or a set of flashcards get a usable answer far more often. • If an app asks the model on your phone about a photo, it now says plainly that it can't read images and points you to a server with a vision model, instead of guessing. Everything on your device stays on your device.

On-device chat that stays quick and cool. • A faster engine. Models on your phone used to re-read the whole conversation before every reply, so each message took longer than the last and your phone got warm. Now they pick up where they left off and read only what's new. • Long chats that don't fall over. Conversations are fitted to the model's real memory, setting the oldest messages aside in one step when a chat gets long, instead of overflowing. • Lighter on memory. The model's working memory takes about half the space it did, and starts small: it only grows when a conversation actually needs the room. • Kinder to a warm phone. When your phone gets hot, replies slow down to let it cool instead of running flat out, and a small "Cooling down" note appears at the top of the chat. The words are the same, they just arrive more gently. • Switch models from the chat. Tap the model name at the top of the chat to move between Apple Intelligence and your downloaded models. Your choice now sticks between sessions. With thanks to everyone who took the time to tell us what wasn't good enough. Everything on your device stays on your device.

The bigger on-device models now work on the phones they were made for. • Gemma 4 E4B and Qwen 3.5 4B could show "Not enough RAM" on 8 GB iPhones. The check compared your phone against a number slightly above what any 8 GB iPhone reports, so it said no to exactly the phones those models are listed for. It now allows for the memory iOS keeps for itself. • Your model choice sticks. If you picked a model that runs on your phone, MyLLM switched you back to Apple Intelligence every time it started. It now keeps what you chose. • Switch models from the chat. Tap the model name at the top of the chat to move between Apple Intelligence and any model you've downloaded, without going into Settings. With thanks to the readers who wrote in and sent screenshots. Everything on your device stays on your device.

Web search that doesn't trip over itself, and downloads on mobile data that ask properly. • Several searches at once. Some models ask for more than one search in a single step. MyLLM used to send those back with an error, which cost a round and showed you a message about a "missing parameter". It now runs them together and gives the model every result. • Bot checks recognised. When a website answers with a "Just a moment…" check instead of the page, the assistant is told so plainly and looks for another source, instead of trying to read the check page. • Downloading on mobile data. Starting a large model download on mobile data now asks first, in a dialog you can't miss. Choose "Always download on mobile data" to stop being asked, or use the switch on the On-device models screen. Everything on your device stays on your device.

Downloading a model on mobile data works properly. If you're on mobile data, MyLLM checks before pulling down several gigabytes. That check was getting in its own way: it asked you to tap Download a second time, then moved the button out from under you, so there was no way to say yes. • A proper question, with an answer. You now get a straight choice — download now, or wait for Wi-Fi — instead of a line of red text and a button that has wandered off. • Or stop being asked. "Always download on mobile data" turns the check off for good, and there's a switch on the On-device models screen if you change your mind. With thanks to the reader who wrote in about it. Everything on your device stays on your device.

The most capable models yet, on your device. MyLLM's on-device engine has been rebuilt on a current version of llama.cpp — nearly two years of upstream work in one step — and that unlocks a generation of models it simply couldn't read before. • Qwen 3.5, new in the on-device list. The 2B is the best all-round model here for most phones; the 4B is the most capable MyLLM can run on device, for iPhone 15 Pro and newer or an M-series iPad. Both are a hybrid design that keeps a very small memory footprint even in long conversations. • Reasoning stays out of the way. Models that think before they answer now do it quietly — you see the answer, not the working out. • How much memory a model needs is read from the model itself rather than estimated, so a long conversation is offered where it genuinely fits. Photos know where to go • No model that runs on your phone can read an image, but the server you've set up probably can. Attach a photo and MyLLM now asks which you'd rather, and remembers your answer. Settings can change its mind later. • Fixed: attaching a photo could fail with a memory warning that had nothing to do with memory. On-device mode was reading your server's model list as though it were its own. • Fixed: the on-device list could offer models that were never on your device. Ask for one now and MyLLM says so plainly instead of quietly answering as something else. New in the App Gallery • Channels — follow the video creators you actually watch, with no account and nothing tracking what you watch. New videos from everyone you follow arrive in one list, and play inside MyLLM so you keep your place. Everything on your device stays on your device — and now, when something would leave it, you're asked first.

On-device models, fixed. This one is about the models that run on your device, and it comes almost entirely from what you told us. • Watch it answer. On-device replies now appear as they are written. Until now the screen sat completely still until the whole answer was finished — on a larger model, most of a minute of an app that looks like it has crashed. Stop works the whole way through and keeps what has been written. • The app stays usable while it thinks. Generation used to claim nearly every processor core, so the chat would not scroll and the app could be impossible to leave until the reply landed. It now leaves the screen room to run. • Memory is measured, not guessed. How much a model really needs is worked out before it loads. The old guess was wrong in the expensive direction, and it is what left the app frozen and often closed by iOS on the spot. • Short on memory? MyLLM now sets the model down and picks it up again on your next message, rather than being shut down mid-conversation. • Hermes 3 3B is available at last. The largest built-in model asked for more memory than any 8GB iPhone or iPad reports having, so it sat in the list and could never be chosen. It now runs on every device its description promised: iPhone 15 Pro and newer, and the M-series iPads. • Models you add yourself work properly. Adding your own GGUF now asks how it wants its prompts laid out, instead of running everything on an older format that made good models look much worse than they are. Everything here stays on your device.

Safer, faster, friendlier. • Nothing is lost by accident. Long-press an app to move it to the Bin instead of deleting it, and the shared workspace now keeps previous versions of every file it overwrites. Restore from the Bin app, the Shell app, or just ask. • Search inside your files. Ask which note mentions the dentist and MyLLM looks through the contents, not just the names. • Plain-English errors with a Retry button. A rejected key, a rate limit or an unreachable server now says what happened and what to do. • Faster to open, smoother to use. The app no longer reads your whole chat history before showing the first screen, long replies stream without stutter, and widgets are back for everyone. • Code that fixes itself. When a Python run fails, MyLLM reads the error, corrects it and runs again before reporting back. • Plans survive the day. An unfinished checklist is saved with its chat and waits for you. • Bigger text, VoiceOver and Reduce Motion are properly supported. Model downloads resume where they stopped and ask before using mobile data. Everything stays on your device.