• Two new models: a fast chat model that fits in about 1 GB, and a new photo model that reads receipts, flyers and forms. • Updated engine: faster loading and replies on recent iPhones. • Fixes and improvements.

• Ask about your day: once you allow Calendar, Reminders, Contacts or Location in Connect apps, any model can answer from them. • Set alarms, timers, reminders and events from chat. Nothing is added until you tap to confirm. • Your profile is now on every tab. • Morning tips with a ready-to-ask question. • A model that finishes downloading while you're in another app is kept, and you get a notification. • Fixes and improvements.

• New Account screen: your plan, restore purchases and app updates, one tap from chat. • Connect apps: see Calendar, Reminders, Photos, Files and Shortcuts in one place. Everything stays on your iPhone. • Fixes and improvements.

• New models: Bonsai 2 (2B and 4B). Small, fast, and they can read your photos too. • Fixes and improvements.

Ask questions about your photos. • New in Models: MiniCPM-V 4.7 1B (Vision). Add a photo in chat and ask about it — receipts, flyers, forms — all on your iPhone, offline. • Model downloads are more reliable. • Stability improvements.

More reliable upgrade screen. • Prices now load reliably on slow or unstable connections, with a Try again button if they can't. • Stability improvements.

Smart Recall and much faster model loading. • NEW — Smart Recall: ask about something from an earlier chat ("what were those dog names you gave me?") and the assistant finds it. Turn it on in Settings. Runs entirely on your iPhone. • Search your chats by meaning, not just exact words. • Models now start much faster, especially on older iPhones. • No more clipboard prompts when you open the app.

New model: LFM2.5 2.6B. • A fast, capable 2.6B model from Liquid AI, great for writing, summaries and email drafts. Answers start right away. • About 1.7 GB — find it under Models → Liquid AI.

Steadier model loading. • If a model needs safe mode on your iPhone, the app now remembers it and starts there next time — no more repeated failed starts. • Safe-mode loading now tells you it may take up to a minute. • If a download can't finish — for example when your iPhone is low on storage — you now see why.

More devices, more places. • Models now load reliably on iPhone 12 and iPhone 12 Pro. If a model can't start the usual way, the app automatically tries a safer mode instead of giving up. • Much faster, more reliable model downloads in mainland China. • Lists in answers now display properly, one item per line.

A faster start and a steadier app. • A shorter, clearer introduction that shows what the app does, offline, from the first screen. • If a model ever fails to load, the app now tells you and lets you pick another, instead of trying the same one on every launch. • Stability improvements.

A new model and a few fixes. • New: MiniCPM5 2B — a compact model that ranks among the strongest open models of its size, at 1.6 GB. • Apple Foundation now uses the full context window your device reports, so longer conversations fit on iOS 27. • Clearer messages when Restore Purchases can't reach the App Store.

Faster, more reliable local models. • Updated the inference engine. Gemma 4 and Granite models now run with the attention pattern they were trained with, and generation is faster across the board. • iPhone 15 Pro and newer are no longer offered only the smallest starter model — the larger, higher‑quality models are available from setup. • Added support for 12 GB devices. • Restore Purchases now tells you what happened instead of failing silently.

Fixes for reliability and first-run setup. • A paused download now reconnects to the right model when you reopen the app, instead of showing a progress screen that never moves. • If you already have models installed, the app no longer asks you to pick one while it is still loading. • Choosing "Decide later" during setup now installs the recommended starter model, so you land in a working chat.

Small update. Faster to get a first answer, and a few rough edges smoothed out.

New model: Granite 4.2 3B from IBM. Apache 2.0 licensed, and like everything here it runs entirely on your device. It is strong on reasoning and maths for its size. Needs a 6 GB iPhone or better.

New model: Qwen 3.5 2B. A generation newer than Qwen 3 at 1.28 GB, so it fits most iPhones. Apache 2.0 licensed and, like everything here, it runs entirely on your device. Also: models that reason before answering now show a quiet "thinking" indicator while they work, instead of their raw notes.

• Fixed: the app could briefly show the trial-ended screen at launch to customers who had already purchased, before their purchase was confirmed. • The chat screen now says a model is loading instead of showing an empty state while it opens.

• After your first answer, the chat now offers a next step so you can start a fresh task without losing the current one.

• The chat screen now shows real download progress instead of an empty state, so you can see how long the first model takes. • You can type your first message while the model is still downloading. It sends by itself the moment the model is ready.

• Fixed: Workspace could stay locked even after subscribing. If you own Lifetime and subscribed to Workspace, it now unlocks correctly. • Stability improvements.

A new layout: Chat, Models, Workspace and Memory each have their own tab, and the app opens straight into a conversation. • Your first model now downloads automatically while you finish setup — you can start chatting sooner. • New models: Liquid AI's LFM2.5 family, a coding-tuned Qwen, and a compact Qwen 3 4B at two-thirds the size. • The model list is organised by publisher, and every download is now verified and checked against free space before it starts. • Create and play: ask any model for a small HTML game and run it in the built-in preview — touch controls appear automatically, and errors are shown clearly instead of a blank page. • Memory has a new screen — see everything the assistant knows about you, delete anything, and turn automatic saving on or off. • Workspace: connect Calendar and Reminders and get your day summarised entirely on your iPhone. • Your phone slows generation down instead of getting hot, with Fast, Balanced and Keep-cool modes. • Fixes: the upgrade screen could appear in the wrong place, and the keyboard could cover the screen with no way back.

• Shortcuts can now run YOUR model. Every action has a Model option — pick Apple Intelligence for speed, or the model you downloaded in the app. • Shortcuts actions now work on devices without Apple Intelligence instead of showing an error. • The model you select is remembered between launches.

Reliability improvements and a refreshed App Store listing.

NEW: Bonsai 27B (1-bit) — the first 27B-class model to run on your iPhone. Full reasoning power, fully offline, in just 3.8 GB. Also new: • MiniCPM5 1B — a fast, modern starter model (0.7 GB) • Faster response generation • Chat reliability fixes