The newest on-device models, a faster engine, Apple Intelligence, and image input. NEW MODELS Qwen3.5 (0.8B to 9B), LFM2.5 2.6B and 8B-A1B, Nanbeige4.2 3B, MiniCPM5 1B, and Google's official Gemma 4 QAT builds. Qwen3.5 4B is the new recommended default: 262K context, thinking mode, and tool calling. Every download is verified against a pinned digest. APPLE INTELLIGENCE On iOS 26 with Apple Intelligence enabled you can start chatting immediately with Apple's on-device model, no download required. Still fully private. IMAGE INPUT Vision-capable models (Qwen3.5, Gemma 4) can now look at your photos directly instead of reading extracted text. Download "vision support" from a model's detail page. FASTER ENGINE Updated llama.cpp with per-device Metal flash attention, a q8_0 KV cache option that fits twice the context in the same memory, and warm-up on load so the first reply starts sooner. Thinking off now works on every reasoning model. A REFRESHED LOOK A sleeker design with Liquid Glass on iOS 26, a cleaner chat screen, and clearer model cards.

Privacy & stability improvements. All 17 models now have verified integrity checksums — downloads are validated on-device before the model activates. Fixed a bug where network-tool consent state wasn't saved correctly across sessions. Widget now shows the accurate downloaded model count. Improved privacy disclosures throughout the app.

Stability improvements

Version 1.3 — The biggest update yet. NEW MODELS • Qwen 3 (0.6B, 1.7B, 4B) — Alibaba's latest generation with built-in thinking mode. Watch the AI reason through problems step by step before answering. Toggle thinking on or off per message. • Google Gemma 4 E2B — Ultra-efficient 2B model that packs surprising capability into under 1 GB. Our new recommended pick for most users. • Google Gemma 4 E4B — More horsepower when you need it. 4B parameters with native tool calling for agent mode. THINKING MODE Qwen 3 models can now show their reasoning. Tap the brain icon in the toolbar to enable thinking mode — the model works through the problem internally before giving you an answer. Great for math, logic, and complex questions. LIGHT & DARK MODE Settings → Appearance now offers Light, Dark, or follow the system. The default stays dark, but Light mode is finally here for those who prefer it. AGENT MODE (MAJOR UPDATE) The built-in AI agent has been completely overhauled: • 23 native tools — web search, web browsing, OCR, camera, image analysis, code interpreter, PDF reader, document reader, file reader/writer, clipboard, translation, URL dispatch, data analyzer, scratchpad, memory, plan mode, and more • Approval mode — choose whether tools run automatically or need your sign-off first • Verbose mode — see exactly what the agent is doing and why • Max steps cap — set how many tool calls the agent can make before wrapping up • MCP server support — connect to local MCP servers on your home network (Home Assistant, custom tool bridges, etc.) • Agent Live Activity — track agent progress right from your Lock Screen or Dynamic Island TEMPLATES Tap the template icon to jump-start common tasks: • Scan a Receipt — point your camera at a receipt, extract all line items, export to CSV or add totals to Reminders • Business card scanner, photo description, whiteboard capture, plant and food identification, fix my writing, email draft, and more MODEL BROWSER IMPROVEMENTS • Three clear states — Active (loaded and running), Downloaded (on device, not loaded), and Not downloaded • Unload button — free up RAM without closing the app • Swap confirmation — clear dialog when switching models so you never accidentally interrupt a session • Checkmarks now refresh immediately after deleting a model EDIT & EXPORT • Edit any sent message and regenerate the response from that point • Export the full conversation as JSON STABILITY Dozens of fixes across this release, including stop-race guard (no more duplicate responses), Vision OOM prevention for large images, auto-load RAM gate (skips auto-load if available memory is too low), widget sync fixes, keychain credential update, and much more.