The model

Frontier intelligence, on your machine.

Eugeny is a 28 billion parameter model. Its European launch is estimated for December 12. Latent thinking, fast decision heads, and one conditioning stack shared by vision, image, video and voice. Built for speed on Apple silicon, and built to stand with the frontier from the desk in front of you.

See what it can do

Illustration: a woman rests her chin on her hand behind an open laptop at a wooden desk. Beside her are a stack of books, a mug, an open notebook and pencils, with a plant and a tall window behind her showing a town with a domed church, a stone bridge, cypress trees and sea.

Model: 28 billion parameters
Estimated launch: December 12
Senses: Text, vision, image, video, voice
Runs on: Apple silicon

A broader perspective.

One local model, many ways to make your ideas real.

Latent thinkingConsider a question deeply. Decision headsMove quickly from question to action. A shared foundationConnect understanding across media. Made yoursAdapt the model to your work.

Inside the model

Five ideas, one machine.

It thinks before it speaks.

And it does not think out loud.

Most models narrate every step, then bill you for the narration. Eugeny deliberates in a continuous latent space: the reasoning happens before language, the compute goes into the problem, and the answer arrives as prose once the work is done.

Ask for depth and it takes the depth it needs. Ask for a quick answer and it gives you one, without a detour through a transcript nobody asked for.

Speed where speed matters.

Some calls should never cost a paragraph.

Routes, labels, matches and approvals resolve through fast decision heads, including the CLM heads, inside a defined answer space. The result is a decision, ready to act on, not prose that has to be parsed back into one.

The heads are small by construction, so a classification arrives without waiting on a full generation. One call, one answer, no schema and parser stitched around it.

03Conditioning

Every sense, one understanding.

Vision, image, video and voice read the language model's conditioning.

One vocabulary serves the family, and one reading of the request flows into every model that answers. Shared context across media, with no duplicated text encoders to keep in step.

Assembled stacks carry near copies of the same weights and several tokenizers. Eugeny carries one understanding, light enough to run locally and consistent enough to trust across every output.

Text
Vision
Image
Video
Voice

Eyes on the real work.

Vision reads images as tasks, not captions.

Eugeny understands a scene, detects what is in it, segments regions, counts objects, reads documents and describes what it sees, in the same conversation as the text. Point it at a shelf, a label, a page, a frame from a video, and it works.

The task by task detail lives on the capabilities page.

Fast on M5. Yours to tune.

Local speed, local privacy, local control.

Eugeny is engineered for Apple silicon: quick answers on M5, with the model running on the machine in front of you. Finetune it on your own material and it learns your vocabulary, your products and the calls your work turns on, inside your own deployment.

The native app brings conversation, files, images, video and voice together with team knowledge. Soffio shares skills and memory; Matchbox connects services with scoped access.

Illustration: a woman reads a book while sitting on a stone wall beneath an olive tree, above a coastal European town with a domed church, cypresses, hills and sailboats.

The release

Estimated European launch: December 12.

Latent thinking, fast decision heads and one conditioning stack, on the machine in front of you.

The ideas behind it

Keep reading.

Three pages that carry the rest of the story.