Textile Labs: small language models, made from scratch

Small language models, made from scratch. Every Loom is trained by us on open data, runs on your own device, and tells you when it doesn't know.

News

The Loom 4.5 familyComing soon

Seven models from one training run, from 11 million to 130 million parameters. Built to read what you give them, use the tools you list, and say how sure they are.

Loom Planck

Our smallest models yet, from 300 thousand to 1.2 million parameters, for jobs where a decision matters more than a paragraph.

Loom Weave 4

Our best-scoring Loom on the honesty test so far, 127 out of 133, with a chat template built in so it works in phone apps. You can try it in your browser.

What every Loom does

Small on purpose

Megabytes, not gigabytes. They run on a laptop or a phone with nothing sent anywhere.

Honest first

A Loom says when the answer isn't in what it was given, and won't pretend to be anything else.

Ours, end to end

Trained from scratch on openly licensed data. No borrowed weights, MIT licensed.