thinking

Working close to large models made me interested in small ones. Scale works — that is settled, and it is also the least interesting thing about it. The question I keep returning to is the inverse: how much can you remove and still have the thing work?

Small language models, model merging, model breeding. Taking models that already exist and combining them into something narrow that solves one problem well under a real compute budget. Most problems do not need a frontier model. They need a small, well-shaped one — and we are still bad at making those on purpose rather than by accident.

That is what Marin is for. Research and development on SLMs and merging techniques, aimed at niche problems where compute efficiency is the binding constraint rather than an afterthought.

What does a merged model inherit that neither parent had?

Is efficiency a constraint you accept, or a taste you develop?

If a model is small enough to understand, is it still large enough to be useful?