Foundation models such as GPT, Claude, and Gemini exhibit impressive capabilities, but how they are built remains a secret. Open-weight models such as DeepSeek provide only the weights, but not the data or exact training recipe.  Opening up this important technology will provide the transparency required to usher in a new wave of innovation.  

‍

Marin introduces a new paradigm for research and development of foundation models: every experiment, successful or not, is conducted in the open. This includes all discussions, twists, and turns.  In other words, the entire scientific process is laid out for scrutiny, enabling scalable and decentralized collaboration.

‍

Marin shares (i) the infrastructure for running experiments reliably and efficiently, and (ii) the science, that is, the knowledge for how to train models that only exists in frontier labs. Anyone in the world can contribute to and leverage Marin. As an example, Open Athena researchers have already begun to use Marin to train DNA models.

‍