Hal keeps getting more capable — and more honest about how he works. • Ternary Bonsai 8B — Hal's deepest on-device model yet. An 8-billion-parameter reasoner that runs entirely on your iPhone, about as fast as the smaller models. • A privacy lock you can see. A lock in the toolbar shows at a glance whether Hal is in a state where data could leave your device. Tap it to understand why. • Sharper memory. More accurate long-term recall, an optional new embedding model, and rebalanced search so the right memories surface — and recent ones are weighted correctly. • Download models once. Models are shared with our companion app, so you don't re-download. • A more honest Hal. His self-reflections are now written in his own first-person voice, and his Self Model screen reads like plain language instead of raw data. When Hal doesn't have something in memory, he tells you — instead of guessing. • Smoother and steadier: faster model switching and loading, and settings that stick.

Version 2.0 — significant update. - Four curated local AI models now available: Gemma 4 E2B, Llama 3.2 3B, Qwen 3.5 2B, and Dolphin 3.0. Each runs entirely on your device. - Self Model: Hal now builds a structured self-model of you over time from patterns in conversation. Browse and edit it in Settings → Self Model. Each entry is private by default. - Salon Mode: put up to four AI voices in conversation. Independent and context-aware modes. - Per-model settings: temperature, memory depth, RAG configuration saved independently for each model. - Upgraded retrieval (optional): Nomic Embed Text v1.5 available as an opt-in via Model Library — significantly better recall on specific facts than the default. 522 MB. - Per-turn memory pre-flight: when a turn would exceed the active model's working memory, Hal explains why in chat rather than crashing. - Real token streaming: responses stream as they generate. - Background downloads: model downloads continue while the app is backgrounded or the phone is locked. - Unified model status dots: one dot, one meaning across all screens. - Significant performance and stability improvements.

What’s New in Version 1.5 This release brings major improvements to transparency, control, and cross-platform consistency. • Clearer On-Device Privacy Updated AFM vs. Phi-3 behavior with precise explanations of when AFM may use Apple’s Private Cloud Compute and how to ensure 100% on-device operation. • Timestamp-Aware Memory Hal now interprets recency using message timestamps, improving what he remembers and how context is weighed. • Expanded Power User Controls Adjust recency weighting, semantic relevance, memory half-life, RAG retrieval size, and temperature — with changes reflected directly in the chat. • Detailed Token View See a full token breakdown for each message in the detailed view. • Cross-Platform Parity Unified layout and behavior across iOS, iPadOS, and macOS, including a resizable Mac window. • Improved Usability Better keyboard dismissal on iPhone/iPad and clearer document upload status. • Restore Defaults One tap to reset all tuning parameters to the original factory settings.