Changelog

Added

• Ministral 3 3B — new edge-optimized model from Mistral AI with chat, coding, and multilingual capabilities (replaces Mistral 7B — smaller download, runs on more devices)
• Tappable citation links — inline [1], [2] citations in web search responses now open the source URL directly in Safari
• Configurable thinking budget — control how long models spend reasoning (256–4096 tokens) with a new Settings slider
• Continuation generation — when the thinking budget is reached, the model transitions smoothly to producing an answer instead of cutting off
• Download confirmation — new dialog when downloading models that may run slowly on your device

Improved

• Smarter model recommendations — Qwen 3.5 2B is now the default for 8GB devices (better speed/quality tradeoff)
• Leaner model library — removed legacy and redundant models for a focused catalog of 12 models from 5 AI labs
• Memory management — RAM checks now run before loading models, preventing out-of-memory crashes
• Onboarding — sub-8GB devices default to Qwen 3.5 0.8B for a smoother first experience
• Source pills — numbered labels (“1 · example.com”) now map directly to inline citations

Fixed

• Memory errors now surface specific messages instead of generic “generation failed”
• Chat template detection for Ministral model IDs
• Missing translated strings across all 10 non-English languages
• 4B+ models now correctly show “May run slowly” on 8GB iPhones

Added

• Qwen 3.5 model series — 0.8B, 2B, 4B, and 9B variants with next-generation reasoning
• Qwen 3 model series — 0.6B, 1.7B, 4B, and 8B with built-in thinking mode
• Thinking mode toggle — switch between fast responses and chain-of-thought reasoning
• Expanded model library — now 15 models from 5 AI labs

Improved

• Model download reliability and resume support
• Chat response streaming performance
• Memory management for larger models
• Model card UI with capability badges

Fixed

• Occasional crash when switching models mid-conversation
• Voice input not activating on first tap in some cases
• Markdown rendering for nested code blocks

Added

• On-device AI chat powered by Apple’s MLX framework
• 7 AI models at launch — Llama 3.2, Qwen 2.5, Phi-4 Mini, DeepSeek R1, Mistral 7B
• Voice input — on-device speech recognition, no cloud processing
• Text-to-speech output for hands-free use
• Web search via DuckDuckGo — privacy-preserving, no search history
• Projects — organize conversations with custom system prompts
• Persistent memory — context that carries across conversations, stored locally
• Markdown rendering with syntax-highlighted code blocks
• 6 accent color themes with light and dark mode
• Siri Shortcuts integration
• Full offline support — works in airplane mode after model download
• Zero accounts — no sign-up, no API keys, no subscriptions