OpenSMLthe name.
OpenSML brings together Open, for open weights, and SML, for small language models.
Small models leave room to experiment: train them, run them locally, and look closely at how they learn. Sharing the weights and the work behind them gives others a starting point to explore, adapt, and build on.
from the first token.
OpenSML was created to explore the full process of building small language models: preparing data, building tokenizers, training from scratch, evaluating checkpoints, and fine-tuning models.
The work happens on Apple Silicon. Our current setup brings five Macs together for distributed training, with each machine contributing to one shared model.
Inside distributed trainingresearch you can explore.
We share released models, training code, and technical reports so others can explore the work, build on it, and understand both the results and the limitations.
OpenSML-150M is the first public release. Its weights and training materials are available for personal, research, and commercial use under their published license terms.
Explore OpenSML-150M
william zebrowski
founder