AI & ML interests
Small language models, training from scratch, and distributed training on Apple Silicon.
Recent Activity
OpenSML
small models. bigger possibilities.
OpenSML is an independent research lab exploring the full process of building small language models—from preparing data and training tokenizers to pretraining, evaluation, and fine-tuning.
We build from scratch on Apple Silicon. Our current training setup connects five Macs for distributed training.
We share model weights, code, and research findings so others can explore, experiment, and build on the work.
Latest release
Solstice-1-175M is a 175M-parameter English chat model. We pretrained it on 42B tokens across five Macs, then fine-tuned it for short conversations, reading comprehension, and instruction following.
Solstice technical report · Source code
First release · OpenSML-150M · 150M technical report
Contact: william.zebrowski@me.com