kimeru

kimeru (Japanese 決める, "to decide") is xvoidsx's planned decision model — a small, fast, local-first judge for agentic software, built to own the decision layer of the stack rather than rent it.

Status: planned. No weights have been trained yet. This card stakes the project and its design. Weights will land here when they exist, and this card will say so plainly when they do.

What it is

A System One-style decision model: it takes a state (text, a ticket, a snapshot, any JSON) plus typed questions and returns calibrated probabilities in a single forward pass, with zero output tokens. It never generates text. The three question shapes:

Choice — pick one of your named options, with a probability for each.
Noul — a yes/no question, answered with a probability between 0 and 1.
Score — a rating on an ordered scale, answered with a weighted position.

It serves over the /v1/systemone endpoint, so any Jev-compatible client can use it unchanged.

Why: permacomputing first

kimeru is designed to be usable on the hardware people actually own — including old machines. A 0.8B to 4B judge that runs locally in milliseconds, with no per-call fee and nothing leaving the machine, fits the same computers navi supports, where a 27B gatekeeper needing datacentre GPUs never would. If intelligence is going to live on your own computer, the decision layer has to live there too, or the whole thing quietly depends on someone else's cloud again.

That constraint shapes every other choice: small size, local serving, and narrow, well-defined jobs over general ones.

Scope: narrow first

kimeru's first jobs will be a small set of well-defined decision tasks inside agentic software — routing, gating, and picking between enumerated options — chosen as the research narrows them down. It will not start as a general decision model.

A decision model cannot drive an open agent loop on its own: it cannot generate text or tool calls. That job belongs to a generative chat model. kimeru is the picker and the gate next to it, not the driver.

Recipe

The recipe is public and small, following the path proven by the open decision models below:

A small open base plus a pointer head, LoRA-trained on labeled state + question + answer data.
Temperature fitted once on a held-out split, so the probabilities are calibrated, not just confident.
Apache 2.0 all the way down — base, head, and weights.

Data

kimeru's training data will be drawn from xvoidsx's own recorded agentic work, chosen as the research narrows the first applications. Nobody else has that data. That is what makes it ours.

Serving

Ollama (/v1/systemone) or llama-server with GGUF. Local-first: the model runs where the agent runs.

Eval before trust

No threshold ships on vendor numbers. Before any threshold is trusted, kimeru gets its own benchmark of 50+ of our own examples, with neutrally named options (option wording steers answers) and a confidence band that escalates unclear inputs to a human. Small decision models are known to get overconfident exactly on the inputs a person should check, and to weaken markedly outside what they trained on — the 4B class is the sane default size, and narrow scope is the honest one.

References

kev (Jared Palmer) — the open reference implementation, including its --init_from path for adapting a checkpoint in a few hundred labeled examples.
tev1 (Together AI) — including its train-your-own recipe.
d1 (Liquid AI) — the category reference.
OpenJev — an open decision model in the same interface.

xxvoidsxx
/

kimeru