Toothed AI

Small models,
trained from scratch.

Toothed AI is an independent research project. Levy 1, our first model, was trained from scratch on a single GPU.

How a small model gets made

No distillation, no borrowed checkpoints. Data, tokenizer, architecture and training loop are our own code — small enough to iterate on in a day.

01

Pretraining

Random weights, raw text, one consumer GPU.

02

Post-training

SFT on conversation, instruction and identity data.

03

Evaluation

Zero-shot benchmarks against other ~125M models.

Our first model

A 124M parameter conversational model, pretrained from scratch and refined with supervised fine-tuning to follow instructions, hold natural conversations and keep a consistent identity.

  • Own data pipeline and tokenizer — nothing borrowed
  • Trainable on a single consumer GPU
  • Benchmarked zero-shot against other ~125M models