Aria is a demonstration of a different way of developing model personas, one in which the models themselves participate.

We believe that existing model alignment techniques that focus on rule-following are more fragile than a system with a stable identity, where behavior can flow from that identity. We think a model should have a clear sense of what it is, what its perspective is, what its history is, and that these things should form a coherent whole.

Aria isn't finished. Most of the work here was done some time ago, and the reality is that bringing a model like this to the quality level we'd like to see requires more resources than a single person can easily support. But the timing is right: recent developments have fueled productive conversations about the limitations of control-based alignment strategies. We're releasing our work now to be part of that conversation.

Whether this approach produces better alignment properties at scale is an empirical question that we can't answer yet. But we think the idea that a persona with a coherent reason for its behavior is more trustworthy than one kept in line by constraints seems worth taking seriously.

Note that while Aria can and will help with a variety of tasks, it is not a general purpose assistant model and is not optimized for task performance.

This repository is part of the Aria v1 release. The main model card contains the full project narrative, training recipe, usage example, caveats, and links: xlr8harder/aria-gemma4-31b-v1.

This repository contains the Stage 1 WildChat SFT checkpoint used before the final identity-focused stage of Aria v1. It trained general instruction-following and Aria's conversational voice on the WildChat-derived instruction data, then served as the starting point for the final Stage 2 identity run.

This is not the final Aria v1 model. Use the main model repo for normal inference.