About Toothed AI
An independent research project exploring small, efficient language models — built from scratch on a single GPU, then measured honestly against their peers.
The project
Toothed AI is an independent research project. The question it tries to answer is simple: how much capability can come out of a compact architecture and carefully chosen data, rather than a bigger run.
That means working small on purpose: models that fit into budgets one person can afford, codebases small enough to understand fully, and experiments that finish in a day instead of a quarter.
Why small models
Bigger runs get bigger models, but they also get bigger costs, bigger teams and bigger black boxes. Small models trade raw scale for access:
Efficiency — less compute, less memory, usable on everyday hardware.
Transparency — every part is small enough to inspect and change.
Reproducibility — a single consumer GPU is enough to train the model.
Focus — tight budgets force better data and architecture decisions.
The approach
No distillation, no borrowed checkpoints. Data, tokenizer, architecture and training loop are our own code — small enough to iterate on in a day.
Levy 1
Levy 1 is the first model to come out of the project: a 124M parameter conversational model, pretrained from scratch and refined with supervised fine-tuning.
From scratch — no distillation, no teacher model.
Conversational — follows instructions and holds natural conversation.
Consistent identity — refined to stay in character as Levy 1.
Benchmarked — zero-shot, against other ~125M models.
Principles
This site
Everything on this site — fonts, icons and the chat page — is served from our own server. The chat is free and requires a simple account with email verification; it sets one essential login cookie and no tracking. Chat messages are processed in memory and are not stored.
Contact
Questions, ideas or feedback? Use the chat on this site — it is the fastest way to reach the person behind the project.