Hikaru-1 enters development.
Work begins on our first model, a finetune of IBM Granite 4.1 (8B)
trained with Explanatory Reinforcement Learning. The first concrete test
of our central research direction.
Update, September 27, 2026: Hikaru’s base changed from IBM Granite to Ministral-3 (8B), and our ERL implementation now uses explanations to guide retries. See the current Hikaru-1 page and ERL research notes. This post otherwise records the work at its original publication date.
Sential Innovations has begun development of Hikaru-1,
its first model. Hikaru-1 is a finetune of IBM Granite 4.1 (8B),
trained using our research method, Explanatory Reinforcement Learning
(ERL).
The decision to build Hikaru-1 answers a single question: does a model trained with explanatory reinforcement learning behave measurably differently from one trained with a scalar reward baseline? Until that question has a measured answer, larger ambitions would be built on assumption. Hikaru-1 exists to remove the assumption.
Why a finetune
Hikaru-1 is not a model trained from scratch. It begins from
IBM Granite 4.1 (8B) and applies the ERL training stage on top.
This keeps the experiment focused on what ERL changes, rather than
entangling results with the cost and variance of pretraining. A
manageable base lets us run, measure, and iterate quickly.
What ERL changes
Conventional reinforcement learning can use a scalar reward, which may be binary or graded. Our ERL experiments add explanations and separate feedback dimensions, using:
- A graded scale from
−5to+5, with0for neutral, so the model learns magnitude and direction, not only verdict. - Per-dimension scores, so a response can be helpful but cold, or accurate but evasive, and be corrected on each independently.
- A grader explanation written for every score, the reasoning behind the correction, which the model learns alongside the score itself.
The hypothesis is that a model trained this way learns transferable principles rather than narrow patterns. Hikaru-1 is the first chance to see whether that holds.
What happens next
Development is underway. The near-term plan is to complete the ERL training run on Hikaru-1, then evaluate it against a baseline trained with conventional RLHF. Results, when we have them and are confident in them, will be published on our research page.
We will share more when there is something concrete to report. Not before.