Sential Innovations
Building the models and tools that make AI understandable.
We are the technology division of Sential, a community driven by curiosity. We build AI models that learn from explanation rather than reward alone, and the software and hardware to put them to work.
Introducing the Hikaru Series
In development
What we are working on
A family of model series, each built for a distinct job. We show their status honestly. Only Hikaru is in development; the rest are still concepts.
Hikaru-1
Our first model, a finetune of IBM Granite 4.1 (8B) trained with our ERL method. The testbed for a new way to make reinforcement learning teachable.
Read more →Arcus
A large model built specifically for software. Notes and plans only; no development has begun. The aim is depth in code, not breadth.
Read more →Nero
Built for our Verplex glasses and for robotics. Combines vision, voice, and language into a single model that understands the physical world.
Read more →Deriva
A general-purpose model with vision. Capable across a wide range of tasks, including reasonable competence at code, but with no single specialty.
Read more →Hirose
A powerful creative generalist, built to help make things across writing and art, but never to replace the people who do that work.
Read more →OBIKI
Our flagship model: a massive, fully custom system we intend to be genuinely frontier, aimed at coding, bioscience, and cybersecurity.
Read more →Yuki
The smallest, most efficient model we can build, made to punch far beyond its weight. Yuki means snow in Japanese.
Read more →PICO
A speech-to-speech model, voice in and voice out with no text in between. One of our more developed concepts.
Read more →Photon
Our image-generation model. Photon-1 turns written prompts into images. A concept for now.
Read more →Products
Kova: one app for any model.
Kova is an inference application that runs our models and any other provider you connect with an API key. One place to compare, chain, and build with any model.
Kova is not yet available. Follow us on social media or check back here for launch updates.
Research
Reinforcement learning that explains itself.
Standard reinforcement learning gives a model a binary signal: good or bad. Our method, Explanatory Reinforcement Learning (ERL), replaces that with an eleven-point scale and a grader that explains its reasoning.
The result is a model that learns why it was rewarded, not just that it was.