MARI — On-Device AI

MARI runs language models on your iPad. Not in the cloud — on the device in your hands. You pick a model. MARI downloads it once. From then on it answers with the iPad's own chip, whether you're on a plane, in a basement, or just tired of sending everything you type to someone else's server. WHAT IT DOES • Chat with open models from Qwen, Google, Meta, Mistral, IBM, Microsoft and others — 32 to choose from, with new ones added to the list without an app update. • Bring your own. If you already have a model folder, import it and MARI will run it with the same engine. • See what each model costs before you commit: download size, memory it needs, how heavy it is on your iPad, and how fast it actually ran last time. • Tune how it talks. Creativity, word choices, repetition, context length — each one explained in plain words, right next to the slider. • Keep going. Start a new chat while another one is still answering. WHAT IT DOESN'T DO • No account. There is nothing to sign up for. • No cloud. Your questions and MARI's answers stay on the iPad. The app has no analytics and no tracking code. • No internet, once a model is downloaded. Airplane mode works. THINGS TO KNOW • Models are downloaded from Hugging Face. Small ones are under 1 GB; the largest is about 8 GB. • Answers come from the model you chose, and small models make mistakes. MARI shows you which model wrote each answer so you can judge for yourself. • Bigger models need more memory. MARI tells you when one is likely to be too heavy for your iPad, but the choice is yours. • Once a day, MARI checks our server for the current model list. It sends nothing about you. Support messages go out only when you send them. MARI is made by one person and runs on Apple's MLX framework. If something breaks, tell us from Settings → Support. We read every message. Requires an iPad with an Apple silicon chip.