A small research lab, studying how AI actually works.
Zellige AI is an independent research lab exploring new approaches to intelligence and publishing everything we learn. We study how frontier models represent knowledge, how information flows through them, and what makes certain architectures work better than others.
From research to models
When our research leads somewhere useful, we turn it into models. Mosaic Pro and Mosaic Micro are our first two, both still in development. We plan to release open weights and offer chat access through a waitlist when they're ready. The research comes first; the models follow from it.
How we work
Three principles shape what we do and what we don't.
Open research
Every experiment gets documented and published, including the ones that fail. The field moves faster when people can build on each other's work.
Mechanistic rigor
We study how models actually work: attention patterns, knowledge representations, information flow. No benchmark theatrics. We care about what holds up under scrutiny.
Experimental
We try approaches that haven't been widely studied. Sometimes they don't work. When they do, they tend to be the things that matter.
Team
We're a small team building in public. Team pages and public hires are coming. Reach out through our contact page if you want to collaborate.