Releases04 Sep 20265 min read

Kaer-R1-mini 1.5B — release notes

A 1.5B distillation of the Kaer-R1 halting behaviour for edge and high-volume serving. What survived the shrink, what did not, and the one result that surprised us.

Man in an emerald doublet holding a smartphone

Kaer-R1-mini is a 1.5B distillation of Kaer-R1, built for high-volume serving and hardware that cannot host the 7B. It exists because in the deployments we have instrumented, 60 to 80% of traffic never needed an expensive model — and the thing deciding that split does not itself need to be expensive.

What survived, what did not

MetricKaer-R1 7BKaer-R1-mini 1.5BChange
Hard-reasoning suite61.4%44.8%−16.6 pts
Abstention AUROC (in-dist.)0.830.79−0.04
Abstention AUROC (paraphrased)0.680.61−0.07
Mean thinking tokens790610−23%
Cost / 1k items£3.10£0.74−76%

The result we did not predict

Capability fell 16.6 points. Abstention AUROC fell 0.04. Those numbers are not proportional and we expected them to be — the working assumption going in was that a model which reasons worse would also judge its own reasoning worse, roughly in step.

The tentative reading is that estimating "will more thinking help here" depends on coarse features of the problem that survive distillation, while producing the correct answer depends on fine ones that do not. We are not confident in that explanation and we are not going to dress it up as a finding. What we can say operationally is that mini is a better router than it is a reasoner, and that is the shape to build around.

The pattern this is for

  1. Mini scores every incoming request and answers the ones it is confident about.
  2. Everything below the halting threshold escalates — to Kaer-R1, to a frontier model, or to a person.
  3. You tune one number, the threshold, against your own traffic and your own cost of a wrong answer.

The threshold is the whole design decision. We wrote a guide for picking it.

Limitations

Weights and evaluation card under staged access: [email protected].

← All posts How to run Kaer-R1 →

Common
questions

Why release a 1.5B model at all?

Volume and latency. At high request rates the serving cost of a 7B becomes the dominant line item, and some deployments cannot use a GPU at all. A 1.5B that keeps most of the routing behaviour is useful even when its own answers are weaker, because its main job becomes deciding what to escalate.

Does the halting head survive distillation?

Partially, and this was the surprise. Raw capability fell as expected. Abstention AUROC fell much less than capability did — 0.83 to 0.79 in-distribution. Knowing that you do not know appears to be a cheaper capability than knowing, which we did not predict and cannot yet explain.

Should I use mini instead of the 7B?

Use mini as a router in front of something larger, not as a replacement. Its answers are meaningfully worse; its escalation decisions are nearly as good. That is a specific and useful shape, and it is not the shape of a general-purpose small model.