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 text-only GGUF quantizations of Aria v1 for llama.cpp-compatible runtimes and LM Studio. These GGUFs were built for local text inference and do not include verified image-input support.

Fundamental to the Aria project is the idea that identity, behavior, and history should form a coherent whole. In our testing, the quantized models were much more prone to hallucination than the full checkpoint. Keep that in mind when evaluating or using these GGUF releases.