Marin is an open-source framework for the research and development of foundation models.
Marin addresses the challenge of building large language models by providing infrastructure and methodology for every stage of the process: data curation, transformation, filtering, tokenization, pretraining, posttraining, and evaluation. The project's defining approach is radical transparency—it documents processes, experiments, and decisions as they happen, including failed attempts. This commitment to open development means that all process knowledge required to build foundation models is shared publicly, not just the final artifacts.
Marin suits researchers and organizations building large language models who value reproducibility and learning from the full development journey. The framework has proven flexible enough to support work beyond text, including audio-text models, DNA models, and protein models when used as a library. The project's current focus is pretraining a large mixture-of-experts model with over 500 billion parameters. For those interested in scaling laws and model recipes, Marin's Delphi scaling suite provides a complete reference implementation spanning from 3e18 to 1e23 FLOPs, including released checkpoints, training pipelines, recipe code, development methodology documentation, and plot-ready data.
Development activity shows sustained focus on frontier model training with active work on mixture-of-experts architectures and scaling research. The project maintains comprehensive documentation and makes intermediate artifacts available throughout the research process rather than only at completion. Contributions extend beyond the core team through library usage in external experiments, indicating the framework's utility as a foundation for specialized model variants.
