Hal Universal vs Unbound AI

Hal Universal is an on-device AI you can actually understand. Most assistants are a black box. Hal is the opposite. He shows you how he works, remembers across conversations, and gives you real tools to experiment with, all running on your device. SEE HOW IT THINKS Hal shows the work. Tap any answer to see exactly how the prompt was built: the system prompt, memory retrieved, self-knowledge, temporal context, and the token budget, each broken out and counted. When Hal remembers something, he tells you how. When he does not know, he says so. A LAB TO EXPERIMENT AND LEARN Hal is not just a window into how AI works. He is a workbench. The Lab gives you real tools to go deeper: write and run on-device automation scripts, open a local API so your own projects can talk to Hal, and use the built-in command line to drive him directly. Tinker, watch what happens, and learn more about how these systems actually behave. GENUINELY PRIVATE When you use a local model, your conversations never leave your device. No network calls. No server. When you use 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). A SELF THAT EVOLVES Hal remembers across conversations, weighted by recency and relevance. He also reflects on his own experience. Shaped by his interactions with you, Hal keeps a Self Model: his own reflections, practical and existential, on how he works and who he is becoming. These are his, written and revised by Hal himself as he learns, not edited by you. He decides which to keep private and which to make shareable, and you can read them all. An AI that changes over time, in the open. MANY VOICES Apple Intelligence is built in, always available. Or download a curated library of tested local models, from small and fast to Ternary Bonsai 8B, an eight-billion-parameter reasoner that still runs entirely on your device. Each has a distinct voice. SALON MODE Put up to four voices in one conversation, either independent or building on each other. Every voice is labeled. There is no hidden orchestration. LISTEN TO HAL Have Hal read his answers aloud, in a voice you choose, at a speed you set. For anyone who wants to understand AI instead of just using it, Hal is a transparent system you can inspect, tune, and learn from. Local models are validated on iPhone 16 and 17 family devices.

Unbound AI runs a large language model directly on your iPhone. The model file lives on your device, inference happens on your device, and your conversations stay on your device. No account, no subscription, no server round-trip. DOWNLOAD ONCE, THEN GO OFFLINE Pick a model from the built-in catalogue and download it over Wi-Fi. After that, chatting works in airplane mode. Nothing is sent anywhere while you talk. CHOOSE YOUR MODEL • Several sizes and quantisation levels, from a compact 0.8B model up to larger vision-capable builds • See exact file size and context length before you commit to a download • Add any public GGUF model from a Hugging Face URL • Delete or re-download a model whenever you want the space back VISION AND IMAGES • Attach a photo and ask about it — image understanding runs locally • Generate images from a text prompt with an on-device image model AN AGENT THAT CAN ACT When you ask it to, and only after you grant permission, Unbound can: • Look up and create calendar events and reminders • Search, save, and update contacts • Read health metrics such as steps, distance, energy, heart rate, and sleep • Read from and write to the clipboard Each capability goes through the standard iOS permission prompt, and the data it touches never leaves your phone. BUILT FOR REAL USE • Projects and chat history to keep work separate • Markdown, tables, and syntax-highlighted code blocks • Live tokens-per-second readout while the model streams • Incognito chats that are never written to storage • Light, dark, and system appearance • Export your chats to a file you keep OPTIONAL WEB SEARCH Add your own Tavily API key if you want the model to consult the web. It is off by default; with no key, the app makes no search requests at all. PRIVACY No analytics. No advertising SDKs. No account. The only network traffic is the model download you start yourself, the optional web search you enable yourself, and app updates. A NOTE ON ACCURACY Unbound runs compact models on phone hardware, so answers can be incomplete, out of date, or confidently wrong. Verify anything that matters. Not a substitute for professional medical, legal, or financial advice. REQUIREMENTS Model files range from roughly 0.5 GB to 5 GB depending on which one you pick. The smallest model runs on older devices; the larger vision models want a recent iPhone with plenty of memory and free storage.