Lemma is a model for researchers.
It's trained on a wide variety of research data to be creative, ingenious, and capable of meaningful computational research. Lemma is currently available in research preview.
Research is a creative act.
Good research depends on knowing what to try next given what you know so far: noticing where models fail, questioning assumptions, and connecting earlier results to a new hypothesis. What distinguishes a great researcher from a good one are ideas: both unexpected and grounded in what has already been learned. Current frontier models can carry out increasingly complex experiments, but they still struggle with coming up with these creative ideas.
We’re building Lemma to learn research judgment. Our core bet is that research intuition is a skill that transfers across domains; towadrs this end, we train on diverse computational research problems, rewarding models for proposing distinct approaches, choosing promising experiments, and using the results to decide what to do next. By learning from research across many fields, we aim to build a model that can bring a broader repertoire of ideas to each new problem.
Our research directions
Creative exploration
How can we train models to pursue distinct, promising ideas and learn when to change direction?
Research over long horizons
How can models build on earlier findings, maintain coherent hypotheses, and choose useful experiments throughout an extended investigation?
Intuition across domains
Which research skills transfer between fields, and how can training across diverse problems strengthen that transfer?
Models and tools built together
We develop our models alongside Qualia, our research agent. Working with researchers helps us understand where existing systems struggle and which capabilities would make the greatest difference in their work.
Build the future of research with us
We’re hiring engineers and researchers to build the next generation of research models and tools. If you share our ambition, we’d love to hear from you.