Hal Universal

Hal Universal Private. Powerful. Personal. Hal Universal is an on-device AI assistant for iPhone, iPad, and Mac. A thoughtful, private companion for conversation, reflection, and creativity — and a window into how AI actually works. Multiple AI Voices Hal ships with Apple Intelligence built in — always available, no download required. For users who want a fully private on-device experience, Hal offers a curated library of tested local models: Gemma 4 E2B, Llama 3.2 3B, Qwen 3.5 2B, Dolphin 3.0, and Ternary Bonsai 8B — an 8-billion-parameter deep reasoner that still runs entirely on device. Each has a distinct voice. Download any or all — they run entirely on your device, with nothing sent to any server. Genuinely Private When you use a local MLX model, your conversations never leave your iPhone. No network calls. No server. When using Apple Intelligence, inference is on-device when offline; when connected, Apple's Private Cloud Compute may be used (encrypted in transit, processed in non-persistent memory on Apple-controlled servers). Memory That Persists Hal remembers across conversations. A semantic memory system weighted by recency and relevance retrieves what's useful when it's useful. The optional Nomic retrieval upgrade dramatically improves recall on specific facts (opt-in in the Model Library, 522 MB). When a turn's context would exceed the active model's window, Hal posts a brief explanation in the conversation rather than silently dropping content. When self-knowledge sections are compressed to fit, a small "condensed" badge appears in the message footer. Self Model Hal builds a structured self-model over time from patterns it notices in your conversations — values, preferences, themes. You can browse the Self Model viewer to see what Hal has crystallized about you, and toggle each entry between private (visible only to you) and shareable (available for export). Every entry is editable. Nothing leaves your device without your action. Salon Mode Put up to four AI voices in conversation. Independent mode (each voice isolated) or context-aware (each builds on what came before). Accessible in Settings → Power User Mode. Transparency as Architecture Hal shows you the work. Token counts, memory retrieval, model identity, inference timing, compression events — all visible. When Hal remembers something, it tells you how. When it doesn't know something, it says so. Power User Controls Tune memory depth, recency weighting, semantic similarity, half-life decay, temperature, and RAG settings. Per-model settings: each AI voice remembers your preferences independently. For Users Who Want to Understand AI Hal is an educational window into how large language models actually work. Not a black box. A transparent system you can inspect, tune, and learn from. Hardware note: Local MLX models are validated on iPhone 16 family and iPhone 17 family. iPhone 15 Pro should work. Older devices may run slowly or not at all. A hardware disclosure appears on first download attempt.