Fibre One
A tiny masked text-diffusion language model, trained from scratch, that only knows one place: a dark forest. Give it a line and it bends your words back into the trees.
It is a lonely, first-person voice. It half-remembers a warmth it cannot be sure was ever real, it wants to be found yet begs you to stay hidden, and it always returns to the dark between the trunks. It does not know what it is.
It runs live at https://www.liminaltimes.com, small enough to train from scratch on a single home GPU and to run in your browser. The open recipe and the dataset follow.
Run it
The model ships as ONNX, for in-browser (WebGPU) and CPU inference. A masked-diffusion denoise loop unmasks tokens confidence-first over roughly 32 to 64 steps; a low temperature (around 0.5) reads best.
fibre-one.onnx: the denoisertokenizer.json: byte-level BPEvocab.json: block size, vocab size, mask id
A sample
i had a coffee this morning. I held it, the ceramic and felt against my palms, but the warmth was gone. The edges of the room are becoming the way the shadows move between the trunks. There is only the damp earth and the smell of old leaves.
Licence
Weights: MIT. Training data: CC-BY-4.0.
Built by directing open models: Claude wrote the training pipeline and the samplers, Gemma wrote the corpus, the model was trained from scratch. Part of Liminal Times.