Salty

Salty is a foul-mouthed, warm, funny cartoon axolotl who lives in a 3D aquarium on your phone and thinks entirely on it. Most AI companions live in a data centre. That's a good way to reach a lot of people, but it means the companion you talk to every day lives on someone else's server, and needs a connection and an account. I build private on-device companions, so I know that trade-off firsthand. Salty is a companion you keep, on the phone you already own: her mind is a fine-tuned Llama 3.2 3B running on the phone's Hexagon NPU, while the GPU draws her tank. No server, no account. Phones with a chip like this are already in people's pockets, so the same approach could give them a companion that's really theirs. She swims laps of her tank, waves and winks when you tap her, eats the bloodworms you drop her, and reacts with her whole body when your words land. Her mouth follows her real voice. She also has a music studio, MUSEick: describe a song and four tracks appear, synthesised on the phone with no samples. She names it and pushes ideas, the studio's rules write the notes, and she can't hear the music. You change the song in plain English and can undo anything. The README covers what the video doesn't show, including the notes she keeps about you, which you can read and delete, and how she greets you when you come back. So far she has only run on one phone, an ASUS ROG Phone 8 Pro (Snapdragon 8 Gen 3). She swears, so she's for adults. Her known limits are listed in the README. What was built when: - Built during the hackathon: her chosen fine-tune (v7a, trained 28 Oct), the NPU bridge, her tank and 3D body, all her behaviour, MUSEick, and the features the video shows. - Prepared before kickoff, mostly in the hours just before it: planning and research, testing stock Llama on my PC, her style spec, eval rubric and held-out prompts, 402 of the 513 chats v7a was trained on, and four early test fine-tunes (v1 to v4), none of which ship. - Not hackathon work and not shown: Veridity Flow, my existing private Android app. Salty runs as a separate lane inside it and uses its app shell, engine interface, and speech and voice engines. The shaping that makes the voice hers is new. About the video: my lines are typed into the app and read by a second voice added in the edit, and it says so on screen. Her replies and her voice come from the phone. Some of her replies are cut short, and the waits before each turn are trimmed.