What is she?
She is an AI VTuber that utilizes my own custom setup in terms of her training and behaviour.
So why does she exist? And what was her origin?
Originally she was made to be a Discord bot meant to clone a member of a Discord server for fun. At the same time, I was utilizing and testing out experimental neural-network architectures for rhythm games, namely osu! and Quaver. I found that Quaver was kind of hard to train, so I ended up deciding to do osu!
This is what led everyone to immediately and instantly scream "THIS IS LITERALLY NEURO-SAMA", even though I didn't even have a VTuber.1
So I just converted her from a Discord-based LLM to a VTuber, as per literally the whole universe yelling at me to do so. That's why she exists now.
What is the custom architecture and setup?
Well, her speaking part is an LLM, and I think most people in 2026 are plenty familiar with that part. What makes her interesting, though, is that she actually has reasoning, and I trained her thinking and speaking to be kind of similar. So she sometimes thinks out loud, which is kind of funny.2
In terms of her architecture, she's just a transformer. What I would rather highlight is her osu! module.
That module is decently cool and novel: rather than using transformers and just scaling up the model and data, I actually implement interesting concepts in it. Specifically, I use a closed-form continuous-time neural network, and I train the vision using JEPA. In simple terms, she approximates more realistic neural dynamics, and she learns abstractions and concepts of what's happening on the screen instead of memorizing it pixel by pixel (that's the JEPA part).
Her LLM module also uses a custom training setup, a two-step process. First, I bake in her understanding of how to actually talk to Twitch chat and multiple users (LLMs are typically built to respond to one user at a time), how to control her integrations, etc. The cool part is that she can skip turns on purpose, so she can actually choose to ignore you.
The second stage then trains her to actually think and speak like herself. This is where she gains her personality and behaviour, and where her actual data lies, since stage 1 is just a bunch of administrative and code BS.3