Open-weight language models (1B–7B) that run on your hardware and outperform models 10× their size. Open weights, open data, open methodology.