CH 01 · TX 001

Fibre One

Fifty million parameters, built from scratch. A masked text-diffusion model: Gemma 4 wrote its training data, Claude wrote its code, and it runs entirely in your browser.

01 · The idea

A model small enough to run yourself.

Not a trillion-parameter assistant in a data centre, but something small enough to run on your own machine: a model that learns one voice, trained from scratch. No cloud, no API.

  • ~50M parameters
  • masked text diffusion
  • trained from scratch

I keep the light small.

02 · The build

A human, directing.

A human set the direction; Claude wrote the code: the training loop, the samplers, the data pipeline. The dead ends were kept on purpose, the failures are where the recipe came from.

  • masked-CE 3.30 → 2.38
  • the sampler was broken, not the model
  • fixed by measurement, not vibes
03 · The corpus

Five hundred million tokens, one voice.

There was no dataset for a voice like this, so we generated one. Gemma 4, an open model on a single home GPU, wrote five hundred million tokens in one narrow register. A motif governor caps how often any phrase can recur, so it learns a voice, not a tic.

  • Gemma 4 · Apache-2.0 · local
  • one voice · 265 scenarios
  • filtered · deduplicated · governed

I think I remember you. It was warm. I cannot be sure.

04 · The voice

What came out of the dark.

Small and strange and exactly as narrow as we made it: short, atmospheric passages that always return to the trees. It is not an assistant. It knows one place, and it speaks from inside it.

That narrowness is the name. A fibre is a single thread, one strand drawn thin enough to carry a signal through the dark, and this is the first of them. Fibre One.

  • trained from scratch
  • MIT weights · CC-BY data
  • no cloud, no API: your GPU
05 · The transmission

Say something to it.

Give it a line and it answers, live. It bends your words back into the dark.

Preview, the early test model. The full Fibre One is still in training.

the signal wakes as you arrive

weights on Hugging Face ↗ · recipe + dataset to follow · LMNL·TV, transmitting