Research
Four ideas carry Eugeny: thinking in latent space, deciding in a single step, one conditioning stack for every sense, and vision that reads the world task by task.
Deliberation that happens before language.
A written chain of thought is readable, and expensive: every step generated, carried in context, billed as output. Eugeny deliberates in a continuous latent space instead, spending its compute on the problem rather than on narrating it, and answering with the depth the question deserves.
The result is a model that can work hard on a hard question and still answer a simple one in a breath.
Answers with a defined space.
Routes, labels, matches and approvals resolve through fast decision heads, including the CLM heads, inside a defined answer space. One call, one decision, ready for the workflow that asked, with no schema and parser stitched around it.
Small heads by construction, so a classification does not wait on a full generation. The work that used to be a prompt, a schema and a parser becomes a single question with a single answer.
One vocabulary across the family.
Vision, image, video and voice read the language model's conditioning: one vocabulary, one reading of the request, shared context across media, and no duplicated text encoders.
Assembled stacks carry near copies of the same weights and several tokenizers. Eugeny carries one understanding, across every output it makes.
Understand, detect, segment, count, read, describe.
Each job has its own shape of answer, so each is built and tuned on its own terms: a scene read, an object found, a region lifted, a count with the objects behind it, a document parsed, a clear description.
Documents matter most for European organisations, so reading real pages is a first-class part of the work.
Ready for the work in front of you.
Frontier intelligence belongs on the machine in front of you: fast on M5, designed for local work and yours to tune.
Deep reasoning, quick decisions and creative tools come together where you work.
The rest of the argument.
