Improved model management with fewer repeated storage scans and readiness updates. More efficient Lens service updates. Larger tap target for Image Generation options. Refreshed App Store screenshots.

Improved Share extension hand-offs, faster startup, clearer model download and privacy information, and more accessible legal and parental-control screens. This update also improves local data reset and model dependency reliability.

NEW • Attached images now work with every assistant model. Multimodal models read pixels directly; text-only models use the bundled SmolVLM2-500M on-device visual bridge. • Added persistent per-model generation settings, while app defaults remain the starting point. • Added an explicit default assistant model. New chats use it, while each existing conversation remembers its own selected model. • Added whole-conversation sharing with a clean chronological transcript. • Assistant replies now show tokens/second and generation time with clear icons. IMPROVED • Fixed the keyboard-open composer conflict: completed answers no longer show through the input controls, and the text field is clearly separated from the action row. • Fixed conversation model switching reverting to the saved default instead of loading the model selected for that conversation. • Fixed valid community models such as Jackrong Qwen3.5 appearing in the chat picker but missing from the Models tab and Downloaded-only inventory. • Fixed Local API requests from browsers and agent clients getting stuck at “Generating.” Web preflight, streaming event framing, graceful completion, and tool-call decision timing are now handled correctly. • Conversation model switching now shows every valid downloaded assistant and links to Models when none are installed. • Expanded completed-response spacing repair for image descriptions and product details. • Fixed a case where a text-only model could show a blank answer immediately after SmolVLM2 analyzed an attached image. • The composer hides while a reply is generating and returns without covering the completed answer, footer, or quick actions. • Completed replies keep compact typography and clearer Markdown spacing. • Reasoning streams live again while supported models are thinking. • Long tool cards remain bounded and no longer make the conversation disappear. • Conversation context is preserved more reliably across short follow-up replies. Everything continues to run privately on your device.

Turn your iPhone into a private local AI server for computers on your network. NEW • OpenAI-compatible chat completions and model-list endpoints. • Ollama-compatible chat, generate, and model-list endpoints. • A configurable LAN server port with copyable connection URLs. • Bearer-key authentication with secure key storage, reveal, copy, and rotation controls. • Streaming and non-streaming responses using the language model already loaded on your device. • Linux-friendly setup instructions and ready-to-copy curl examples. IMPROVED • The Mac Bridge remains available with its existing encrypted pairing workflow. • Clearer server status, network-address handling, lifecycle behavior, and error messages. The local API server is disabled by default and intended for trusted networks while the app is in the foreground.

Lens now feels more at home on iPhone. NEW • Added support for the Bonsai 8B and 27B EXT series for more capable on-device assistance. • A redesigned Liquid Glass Lens experience keeps the camera front and center. • A compact model picker makes on-device readiness easier to understand. • Ask, Translate, Scan, and Solve now have clearer modes and subtle visual accents. • New capture, scanning, and result animations provide better processing feedback. • Lens answers are easier to read, with improved hierarchy and streamlined actions. • Accessibility improvements include larger controls, Dynamic Type, VoiceOver labels, Reduce Motion, and increased-contrast support. Everything continues to run entirely on your device — no account, no servers.

A fresh, brighter interface for both Light and Dark Mode. • New Liquid Glass surfaces on iOS 26 and later • Cleaner backgrounds with improved contrast and readability • Refined Home, Chat, Voice, and action controls • Polished translucent materials on earlier iOS versions Your private AI workflow remains fully on-device.

More stable, private, and dependable on-device AI. NEW • Safer Wipe All Data — clears saved API keys, Hugging Face token, and Mac Bridge pairings from Keychain. • Stronger Mac Bridge pairing protection and safer web page fetching. • Conversation saves now report failures instead of failing silently. • iCloud sync entitlement and clearer sync error reporting. VOICE • KittenTTS speaks with less lag between sentences. • Apple System Voice prefers Enhanced/Premium voices and avoids the Bluetooth “zombie” route after voice chat. • Kokoro float16 audio decoding fixed (download Kokoro from Models → Catalog to enable). LENS • Fixed a camera vision-encoder crash (SIGSEGV) on newer iPhones by keeping Core ML off the GPU path. Everything still runs entirely on your device — no account, no servers, nothing ever leaves your iPhone.

A cooler, cleaner, more dependable on-device assistant. NEW • Cleaner, more professional interface focused on function over decoration. • Simpler Home, Assistant, Lens, and Settings screens with quieter controls and fewer visual effects. • Chat and camera controls now use flatter, more readable surfaces. RELIABILITY • Reduced heat and battery impact in Lens mode. Visual captions now run as a short follow-up burst instead of continuing indefinitely. • Lower camera preview frame rate and stronger thermal backoff when iPhone reports serious heat. • Live OCR now runs less often and prevents overlapping recognition work. • Visual model responses use tighter token limits for shorter, cooler on-device runs. • Continued model, voice, and stability fixes. Everything still runs entirely on your device — no account, no servers, nothing ever leaves your iPhone.

A cleaner, more professional look — and a more dependable on-device assistant. NEW • Onyx theme — a refined, near-black professional look, now the default. Prefer color? System blue, Rose, Teal and more live in Settings → Appearance, alongside a true-black OLED dark mode. • Redesigned Settings — simpler, more native menus throughout. • On-device error analysis — let your local model diagnose issues right on your iPhone, from Settings → Diagnostics or straight from chat. • Chat history is now one tap away. RELIABILITY • On-device models load dependably, even when the system is low on memory. • Incomplete model installs now repair themselves instead of failing. • Cleaner conversation titles, smoother voice, and many stability fixes. As always, everything runs entirely on your device — no account, no servers, nothing ever leaves your iPhone.

Still no account, and nothing ever leaves your iPhone. This update fixes an image-generation crash, adds new accent colors, and lands dozens of reliability, privacy, and polish fixes. NEW • More accent colors — Teal, Indigo, Coral, and Magenta join the set, now pickable as color swatches in Settings → Appearance. FIXED • Image generation no longer crashes if the screen locks or you switch apps while a model is downloading. • Read-aloud in the Lens (“speak what I see”) now actually speaks. • Returning to the camera after leaving the app no longer shows a black preview; the live caption loop fully stops when you switch tabs. • Voice stays responsive — speech detection no longer blocks the interface and a reply can’t get stuck mid-sentence. • Resumed downloads keep their sign-in and Wi-Fi-only settings; cancelling a download reclaims its disk space. • Faster launch for people with many downloaded models. PRIVACY • “Erase all data” now also clears the on-device search index and in-memory chats — wiped chats can’t reappear in search or sync. • Deleting a chat stays deleted across iCloud sync. • Pairing with a Mac requires an encrypted connection by default, with a clear opt-in for older Mac apps. • The first-launch legal screen now asks you to scroll through each document before accepting. Plus accessibility, contrast, and stability polish throughout.

Improved visuals, smoother UX, and minor bug fixes for a cleaner, more polished experience.

Still no account, and nothing ever leaves your iPhone. This update fixes an image-generation crash, adds new accent colors, and lands dozens of reliability, privacy, and polish fixes. NEW • More accent colors — Teal, Indigo, Coral, and Magenta join the set, now pickable as color swatches in Settings → Appearance. FIXED • Image generation no longer crashes if the screen locks or you switch apps while a model is downloading. • Read-aloud in the Lens (“speak what I see”) now actually speaks. • Returning to the camera after leaving the app no longer shows a black preview; the live caption loop fully stops when you switch tabs. • Voice stays responsive — speech detection no longer blocks the interface and a reply can’t get stuck mid-sentence. • Resumed downloads keep their sign-in and Wi-Fi-only settings; cancelling a download reclaims its disk space. • Faster launch for people with many downloaded models. PRIVACY • “Erase all data” now also clears the on-device search index and in-memory chats — wiped chats can’t reappear in search or sync. • Deleting a chat stays deleted across iCloud sync. • Pairing with a Mac requires an encrypted connection by default, with a clear opt-in for older Mac apps. • The first-launch legal screen now asks you to scroll through each document before accepting. Plus accessibility, contrast, and stability polish throughout.

Still no account, and nothing ever leaves your iPhone. NEW — Code Mode, rebuilt Point your camera at code on a screen, terminal, or editor and tap once. Code Mode now captures a sharp still, reads the code faithfully on-device (operators and indentation preserved — no autocorrect mangling your symbols), and then lets you: • Extract — get the exact code as clean, copyable text • Explain — a plain-English walkthrough of what it does • Review — bugs, security issues, and smells with concrete fixes • Debug — point at an error or stack trace and get the likely cause and a fix The heavy lifting runs on your on-device coding model, so it understands code far better than before. The extracted text is editable, so you can fix the odd misread before asking. Send any capture straight to chat, or save it as a snippet. IMPROVED • A cleaner camera viewfinder with a simple framing guide for code. • Live OCR is now an independent toggle you can flip on in any mode. • Lighter, faster camera tab — we removed per-frame work that's no longer needed.

CodeLens 2.5 builds on 2.4's on-device assistant with a few focused upgrades — still no account, and nothing ever leaves your iPhone. NEW Generate images right in chat Ask the assistant to picture something and the image appears inline in your conversation — no need to switch to the Images tab. It can fold image generation into a multi-step answer alongside web search, your files, and your Knowledge Base. Save to your Knowledge Base from chat Ask the assistant to remember a snippet or a document and it's indexed on-device, ready to retrieve and cite in later questions. IMPROVED Clearer model badges Models now show at-a-glance badges for what each one can do — Tools, Coder, Fast, and Multilingual — so it's easier to pick the right model for the job. Sharper built-in math The assistant's calculator now returns precise decimal results instead of rounding division down to a whole number.

CodeLens 2.4 is a big on-device update: generate images, chat with your own files, reach the app hands-free, and enjoy a smarter, sturdier assistant — all without an account, and nothing ever leaves your iPhone. NEW Image generation, on your device A new Images section in Models generates pictures from a text prompt entirely on-device. Pick SDXL Turbo for speed or Stable Diffusion 2.1 for detail, then share straight from the preview. CodeLens suggests the model best suited to your device. Siri & Shortcuts Ask CodeLens, start a voice chat, or caption what you're looking at — hands-free. Add the actions to Siri, the Shortcuts app, or your Action Button. Instant answers with Apple Intelligence On supported devices, get a quick reply right inside Siri and Shortcuts — no app launch and no download. Knowledge Base — chat with your files Add your own documents and ask questions about them completely offline. CodeLens finds the relevant passages on-device and cites them in its answer. A smarter assistant The assistant can now search the web, read files, and look things up in your Knowledge Base across multiple steps to handle tougher questions. Search your chats from the Home Screen Past conversations are now searchable in Spotlight — tap a result to jump straight back in. IMPROVED More reliable image generation Generating no longer runs out of memory mid-render, so it finishes cleanly instead of quitting. Smarter model handling Models are organized by type — Assistants, Vision, Image generation, and Voice — and image models are recognized correctly whether you're browsing the catalog or importing your own. Fewer crashes on large models Loading a model too big for your device now stops with a clear message instead of crashing, with an adjustable memory-safety setting. New built-in diagnostics capture crashes and a breadcrumb trail to make issues easier to fix. Reclaim wasted space A "Clean up partial downloads" button recovers the gigabytes left behind by cancelled or failed downloads in one tap. More reliable downloads Downloads retry automatically through dropped connections and flaky networks, and a load-timeout watchdog recovers a stuck model load.

CodeLens 2.3 adds on-device image generation and makes model handling smarter and sturdier. NEW Image generation, on your device A new Images section in Models lets you generate pictures from a text prompt entirely on-device — no account, no network, nothing leaves your iPhone. Pick SDXL Turbo for fast results or Stable Diffusion 2.1 for more detail, type a prompt, and go. The app automatically suggests the model best suited to your device, and you can share the result straight from the preview. IMPROVED Importing models actually works Importing a model from Files now detects what kind of model it is, finds the weights even when they're nested in a subfolder, and tells you clearly if a folder isn't a usable model — instead of silently doing nothing. Fewer crashes on large models Loading a model that's too big for your device now stops with a clear message instead of crashing the app. There's a new, adjustable memory-safety setting if you want to push your device harder. Reclaim wasted space Cancelled or failed downloads used to leave gigabytes of hidden files behind. A new "Clean up partial downloads" button shows exactly how much you can recover and frees it in one tap. More reliable downloads Downloads now retry automatically through dropped connections and flaky networks instead of failing outright, and a load-timeout watchdog recovers a stuck model load.

CodeLens 2.1 makes Voice mode work the moment you tap into it. NEW Voice models, in the box Two on-device text-to-speech engines now ship with the app: KittenTTS Nano (8 English voices) and Kokoro-82M (57 voices across 9 languages). Open a voice conversation on a fresh install and the assistant starts speaking back — no downloads, no setup, no network. The Models → Catalog still lets you grab the higher-quality KittenTTS Mini if you want it. FIXES • Fixed a crash in Voice mode the first time the microphone permission was granted. The audio engine could observe a half-initialized audio session and trip an AVFAudio assertion. The engine now surfaces this as a recoverable error you can retry, instead of terminating the app. • Audited the rest of the audio pipeline for the same family of crashes and confirmed every other AVAudioEngine entry point already fails gracefully.

CodeLens 2.0 turns your iPhone into a local-model lab — download, test, compare, and rank on-device LLMs and VLMs, completely offline. Free, no account, nothing leaves the device unless you say so. NEW Model Arena — A/B any two models Send one prompt to two models, read both answers side by side, and pick a winner. Flip on Blind mode to hide the names and randomize which answer is which, so you judge the output, not the brand. Every vote feeds a persistent on-device leaderboard — an Elo rating and win/loss record built for YOUR device and YOUR prompts. Works with built-in, downloaded, and custom models, and labels each result's quantization and runtime so you can pit 4-bit against 8-bit head to head. Quality eval — score quality, not just speed The benchmark tells you how fast a model runs; the new quality eval tells you how good it is. A pack of auto-graded, on-device checks — strict JSON output, instruction-following, arithmetic, formatting, and code — produces a single quality score plus a pass/fail breakdown. No cloud, no judge model, fully deterministic. A refreshed model roster Added the Qwen3-4B "2507" refresh in both Instruct and Thinking flavors, an 8-bit fidelity variant for the highest-memory devices, and Llama 3.2 1B for the lightest ones. Picker rows now carry capability badges — Vision, Thinking, Recommended, Best — alongside the runtime engine, so you can tell what a model does before you load it. Web search from the assistant The on-device assistant can finally use the web when your question needs current information. It asks first (honoring your Web Access setting), runs the search, reads the top results on-device, and answers with citations — then gets out of the way for everything it can already answer offline. POLISH • The visual A/B compare screen gains the same winner voting, Blind mode, and leaderboard as the text arena. • Finer generation control: sampler settings and an optional JSON output mode for the assistant. • Capability and runtime metadata is now a single source of truth, so badges, device-fit filtering, and tool gating all stay in sync. FIXES • Fixed the assistant being unable to actually run a web search it requested. Under the default "ask every time" setting the model's search now pops a one-tap permission prompt and runs, instead of silently erroring out. • Hardened the on-device generation path against unsupported sampler options so model loads and replies stay reliable across the expanded roster.

> Pair On Device LLM with your Mac. > > The new Bridge connection lets the on-device model see what you're working on — your frontmost app, selected text, Xcode build errors, files in an approved workspace — without anything leaving your devices. Type `/mac` in chat to ask questions about what's on your laptop. > > Also in this update: > • Fixed a crash in Lens when switching to FastVLM Visual mode > • Lens model picker moved inline into the top bar — fewer taps to switch > • Whisper voice model now downloads inline from Settings > • Faster, smoother model loading >

New assistant model picker from the top bar for faster switching between available models • Refined Assistant interface with better loading states and cleaner controls • Major design polish for the Models tab, including improved cards, hierarchy, and browsing flow • Better memory handling when switching between assistant and vision models • Stability improvements and bug fixes across model loading and management