Artificial Intelligence
Models that learn from being taught, not from being scored.
We build our models the way we believe they should be built: with signals that explain themselves. What follows is how we think about AI, and the four series we are working toward.
Our approach
A model should know why it was corrected.
Most reinforcement learning tells a model two things: that an answer was good, or that it was bad. That is a thin signal for a system expected to reason. A child who is only told "wrong" learns little; the same is true of a language model.
Our research direction, which we call Explanatory Reinforcement Learning, changes the signal in three ways:
- Scaled rewardAn eleven-point scale (−5 to +5, with 0 for neutral) replaces the binary good/bad signal.
- Per-section scoresPersonality, helpfulness, and other dimensions are scored independently rather than collapsed into one number.
- Grader explanationA strong grader model writes an explanation of every score, where the response succeeded or failed, and why.
The result is a model that learns the reasoning behind a correction, not merely the correction itself. We believe this is a more honest and more effective foundation.
Model series
Nine series, one in development.
Each series is built for a distinct job. Only Hikaru-1 has left the planning stage; the rest remain concepts.
Hikaru Series
Hikaru. The testbed
Our first and only model currently in development. A finetune of IBM Granite 4.1 (8B), built to validate our ERL training method at a manageable scale.
Hikaru exists to answer a single question: does a model trained with explanatory reinforcement learning behave measurably differently than one trained with a binary reward?
Arcus Series
Arcus. Built for software
A large model dedicated to code. The aim is depth: a system that is exceptional at software engineering rather than passable at everything.
Notes and architectural thinking exist, but no development has begun. We would rather build one thing well than many things averagely.
Deriva Series
Deriva. The generalist
A general-purpose model with vision. Capable across a wide range of tasks, including reasonable competence at code, but with no single specialty.
Deriva is the series we expect most people to use most of the time, the dependable everyday model in the lineup.
Nero Series
Nero. For the physical world
Built for our Verplex glasses and for robotics. A single model that unifies vision, voice, and language, with the orchestration needed to act in the real world.
Nero excels at persona, tooling, and orchestration. The things a model needs when it lives in a body or a pair of glasses.
Hirose Series
Hirose. The creative generalist
A powerful generalist built for creative work, with the range to help make real things across writing, art, and ideas.
There is a firm line with Hirose: it is meant to support human creativity, not replace it. If it ever starts taking work from human creators, we retire it, a rule that holds for every model we make.
OBIKI Series
OBIKI. The flagship
Our most ambitious series by far: a massive, fully custom model we intend to be genuinely frontier, for coding, bioscience, and cybersecurity.
OBIKI is the furthest-out concept in our lineup and the north star the rest of the work points toward. Built from the ground up, not finetuned.
Yuki Series
Yuki. Small and efficient
The smallest, leanest model we can build, made to punch far beyond its weight and stay startlingly capable for its size.
An efficiency project we build partly for the joy of it. Yuki (雪) means snow in Japanese: small, light, and everywhere.
PICO Series
PICO. Speech to speech
A speech-to-speech model: voice in, voice out, with no detour through text. One of our more developed concepts.
By skipping the text step, PICO aims for conversation that feels immediate rather than stitched together from separate parts.
Photon Series
Photon. Image generation
Our image-generation model. Photon turns a written prompt into an image; Photon-1 is the first in the series.
Photon is our entry into visual generation, the counterpart to the reasoning models in the rest of the lineup.