Volunteer compute
Lend idle GPUs and CPUs from home, labs or universities. Many small machines, joined together, train one shared model.
VolunteerDistributed computing
Language models, grown together by the people who use them.
Ygaraty trains open language models on computing power donated by volunteers, and gives them back to the community as models that are simple to run and embed.
Open, embeddable models
Compact models released openly, ready to download, run locally and embed in your apps with a few lines of code.
Research in the open
We study distributed training over unreliable networks and efficient language models — and publish everything we learn.
Sovereign training
Sovereign people need sovereign model training — not concentrated in the hands of a few players.
The name
"Ygaraty" comes from Old Tupi and Guarani — a name rooted in the languages of the land, for a project rooted in its community.
How it works
- Step 1
Join
Open a web page. As simple as that.
- Step 2
Train
Your device works on small slices of training, synced with peers worldwide.
- Step 3
Release
Finished models are published for the community, free to use.
Research on distributed training and language models
- — Training across heterogeneous, intermittent volunteer hardware
- — Communication-efficient optimization over the open internet
- — Small, capable models designed for embedding and on-device use
- — Open datasets, code and results for anyone to reproduce
Lend your machine to the commons.
Every contributor brings open language models closer to everyone.
Get in touch