Efficient Pre-training
Scaling laws beyond raw compute. We study data curation, sparse architectures, and curriculum design to train frontier-quality models at a fraction of the cost — and we release the recipes.
View work →Quanta is an independent research lab developing frontier models and releasing them openly — weights, training recipes, and evaluations — so the broader community can study, audit, and improve them.
We believe the most capable models should not be locked behind closed doors. Transparency is not a constraint on progress — it is the mechanism by which progress becomes trustworthy.
Every system we build ships with its weights, data documentation, training code, and a full evaluation suite. We publish negative results alongside the positive ones, because reproducibility is a feature, not an afterthought. Our work spans pre-training efficiency, interpretability, and alignment — connected by a single commitment: the people who use these systems deserve to understand them.
Each direction is led by a small, focused team and judged by what it contributes back to the open ecosystem.
Scaling laws beyond raw compute. We study data curation, sparse architectures, and curriculum design to train frontier-quality models at a fraction of the cost — and we release the recipes.
View work →Reverse-engineering the internal computations of large models. We develop tools to locate circuits, trace features, and explain behaviour — turning black boxes into objects of study.
View work →Methods for making model behaviour predictable, controllable, and robust to misuse. We build open evaluations and red-teaming protocols that anyone can run and extend.
View work →We hire researchers and engineers who care about getting the details right and sharing what they learn. Small teams, long horizons, real ownership — and everything you build sees the light of day.
See open roles →