Fewer words.
Better reasoning.
Kōzu AI is a small applied-research lab. We fine-tune language models to reason more efficiently, and we build the tools and datasets that make it repeatable.
- Founded 2026
- Part of The Kōzu Group
- Models on Hugging Face
What we’ve released so far.
Three models that scale one idea from small to large, and a fourth that reads the reasoning behind another model’s answers.
All projectsWhether it’s a model, a framework or a dataset, we’ve always loved bringing an idea to life.
That love is why we build the projects that let machines create alongside us. We keep pushing for knowledge, and for what’s possible when you’re willing to be creative about it.
About the labHow the work fits together.
The models are the part you can download. Behind each one are the tools that train it and the datasets that shape it.
Datasets
Quark teaches concise reasoning. Photon, still in preparation, hardens models around compliance.
See the datasetsTools
Hadron, Tokamak and Stellarator generate the training signal, compress it, and run the training.
See the toolsModels
Deimos, Europa, Ganymede and Kuiper are what comes out the other end.
See the models
From the notes.
Research updates and shipping logs. We write when we have something to say.
All insightsA Tinker research grant.
Thinking Machines has awarded Kōzu AI a Tinker research grant. What it pays for, and where it points us.
How we distill.
The tools behind the models: Hadron, Tokamak and Stellarator, and why we distill instead of just scaling up.
The model line.
Four models, one idea: reasoning that does more with every token. Deimos, Europa and Ganymede scale it up, and Kuiper reads the result.