About
Textile Labs is a small, independent lab that makes small language models, called Looms.
Every Loom is trained by us from scratch, on openly licensed data, on two borrowed graphics cards in a single session. They are small enough to run on a laptop or a phone, and they are built to be honest: to say when they don't know, to answer from what you give them, and to never pretend to be something they're not.
We publish every model under the MIT licence, with its scores, its misses and the commands to run it.
Why weaving
A language model is woven from many small patterns crossing each other, so we name ours after cloth. Each line is a size of model.
What we believe about small models
Most of the field is racing to make models bigger. We think small models deserve the same care.
Small is a feature
A model that fits in a few megabytes runs on your own device, works offline, answers instantly and keeps your words private. That is worth more than knowing everything.
Honesty before cleverness
A small model will get things wrong. The useful ones say so: they answer from what you give them, admit what they can't know, and can't be talked into pretending.
Read, don't recite
A small model can't hold the world in its weights. It can learn to read: to find the answer in your text or a search result and quote it back.
Made, not borrowed
We train every Loom from scratch, so we know exactly what went into it, and we can tell you.
Measured, not promised
Every claim on this site comes from a test we wrote before training. When a model misses, we say so.
Our goals
Where we're heading.
Small models you can rely on
Looms that read what you paste, use the tools you list, and are right when they say they're sure.
A full family every generation
From a model for your phone to a model for your laptop, all trained together and released together.
Running everywhere
Every Loom works in your browser, in Ollama, LM Studio, Jan, llama.cpp and phone apps, with nothing to set up.
Open, and on the record
Open weights under MIT, and an honest record of what worked and what didn't, so others can build on it.