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.

Smoother scrolling, everywhere.

If scrolling a long chat or your app grid felt like it was stepping rather than gliding, this is the fix. The app was quietly redoing a lot of work on every single frame; now it does it once and keeps it.

• Chats scroll cleanly: the transcript was being redrawn, and every message on screen re-formatted, sixty times a second while you scrolled. Formatting is now done once per message and reused, and the scroll no longer redraws anything it doesn't have to.
• Your apps and the App Gallery too: every icon on the springboard was being redrawn once a second whether it had anything to show or not, and the gallery re-downloaded and re-decoded artwork you'd already scrolled past. Both fixed, so the grid and the gallery move as smoothly as the rest of the app.

No settings to change and nothing new to learn, it should simply feel right.

Everything here stays on your device.

Your workspace grew folders — and apps that open into it.

The shared space your apps read and write is now a real filesystem. Make folders, move things between them, and browse it properly. MyLLM can do all of it too, so you can simply ask it to tidy up.

• Files, with folders — create a folder, open it, and move a file or a whole folder into it. Swipe any row to move, rename or delete. Deleting a folder that still holds files asks first, and tells you how many.
• Ask MyLLM to organise — "put the shopping lists under lists/ and archive last year's notes" now has somewhere real to go.
• Open a file with an app — hand any document straight to an app that can read it.
• Workbench, new in the App Gallery — the app your files open into. It renders Markdown properly, CSV as a real table, JSON, logs and plain text. Choose the typeface, size, line spacing and column width, switch on Paper mode for long reading, and use the formatting toolbar when you edit. Save straight back, or save a copy anywhere in your workspace.
• Live app tiles — an app can now show one thing on its own icon: a streak, a count, a score. The tile turns over to show it, then turns back.
• Swipe to your apps — swipe in from the right edge of a chat to open MyLLMos, and back from the left edge to return. The handle above the keyboard still works exactly as before.
• A Tools screen — see everything MyLLM can do, grouped by what it's for: which tools are switched off, which ones you actually use, and what each one takes.
• Python can open your files — hand real files to Python and it can read them normally, then write its results back where your apps can pick them up.
• Command Center — start with real monitors instead of a demo, keep a history of what changed, and remove anything from the dashboard by touching and holding it.

Fixed
• Copying a long answer could freeze the app and leave the message menu stuck until you force-quit. Long copies now stay on this device; everyday copies still reach your Mac and iPad.
• Files landed where you expected. MyLLM could save into a private area only it can see and still tell you the file was saved — it now knows the difference, and says so plainly if it happens.
• Connected tool servers work properly. If you connect a server with a lot of tools, most of them never reached the assistant. Tools you connected or wrote yourself now come first, and the assistant can find them when it goes looking.

Everything here stays on your device.