Skip to content
Model cards

Model cards

A model card says, in plain words, what one part of Luna’s AI does, what it reads, what it produces, which model it runs on, where it runs, and the limits it follows. There is one card per aspect, one for drafting plans, and one for safety.

The cards

CardWhat it coversBase model
TrainerThe trainer aspect in chatQwen3.6-35B-A3B
NutritionistThe nutritionist aspect in chatQwen3.6-35B-A3B
KeeperThe keeper aspect in chatQwen3.6-35B-A3B
CounselThe counsel aspect in chat (the default)Qwen3.6-35B-A3B
MuseThe muse aspect in chatQwen3.6-35B-A3B
NaturalistThe naturalist aspect in chatQwen3.6-35B-A3B
Plan draftingPlan revisions, meal food choices, and the check before you see themQwen3.6-35B-A3B
SafetyRefusals and referrals across every functionFixed rules, plus Qwen3.6-35B-A3B

What every card has in common

One Luna, one model call. The six aspects are not six separate models. Each chat reply is one call to the same model, which picks the aspect as it starts answering (see Six aspects, one Luna). Each card describes how Luna behaves when that aspect is in front.

The base model. Luna’s text work runs on Qwen3.6-35B-A3B, made by Qwen and released under the Apache-2.0 licence.

Where it runs. Luna’s AI runs server-side, not on the phone. Chat, plans, photo features, speech-to-text and memory run on Runink-operated servers, and no third-party AI service is called. The apps do still contact a few outside services, such as Google Sign-In and, on the phone apps, the phone’s own speech engine, which reads Luna’s replies aloud. The full list is on Privacy and data.

Numbers come from code. Streaks, points, levels, nutrition targets and weekly set counts are computed in code from what you logged. The model never supplies a number you then have to trust.