# Kestrelyn: kestrel-decider

> Small open decision model from Kestrelyn: a LoRA and pointer head on Liquid's LFM2.5-230M-Base, trained with the strands-decider recipe. Answers yes/no, choice and rubric-score questions with per-option probabilities and no generated text; aimed at CPU and edge routing.

- Page: https://systemonemodels.ai/kestrelyn/kestrel-decider
- API: https://api.systemonemodels.ai/v1/models/kestrelyn/kestrel-decider
- Download: `pip install systemonemodels && systemone pull kestrelyn/kestrel-decider`

## Facts

| | |
|---|---|
| Maker | Kestrelyn (https://systemonemodels.ai/kestrelyn) |
| Decides | choice, score, noul |
| Architecture | strands-decider |
| Base model | liquidai/lfm2.5-230m-base |
| Parameters | 230M |
| Licence | LFM Open License v1.0 (commercial use only below US$10M revenue; non-commercial above that) |
| Availability | Open weights |
| Released | 2026-10-03 |
| Latest version | 1.1.0 |

## Reported evaluation

Suite: JevBench v1 public set (231 tasks) at a 3,072-token window, maker's run with the official harness. Numbers are the publisher's own.

- Decision accuracy: 64.5%
- Calibration error (ECE): 0.064
- Median latency: 19 ms
- p95 latency: 126 ms

## Model card

<!-- generated by scripts/seed_catalog.py; edit content/models/catalog.yaml -->

# Kestrel Decider 230M

Small open decision model from Kestrelyn: a LoRA and pointer head on Liquid's LFM2.5-230M-Base, trained with the strands-decider recipe. Answers yes/no, choice and rubric-score questions with per-option probabilities and no generated text; aimed at CPU and edge routing.

The repo holds the rank-128 LoRA and head; the base downloads from LiquidAI at load time. Per-kind temperatures are stored with the model, and `strands-decider serve` exposes /v1/systemone. An independent release, not endorsed by Liquid AI or the strands-decider authors. The LFM Open License caps licensed commercial use at US$10M annual revenue. The maker says it is weak on answer adequacy, multi-hop reasoning and hard judgement, and that MuSiQue regressed in v1.1. On the maker's own run of JevBench's public items (a third-party benchmark, not submitted to its board) it answers 149 of 231 with ECE 0.064; on 6,000 unseen classification rows it scores 0.558 with ECE 0.058.

## What it decides

- **choice** — picks one option from a set
- **score** — places the input on an ordered scale
- **noul** — answers a yes/no question with one probability

## At a glance

| | |
|---|---|
| Parameters | 230M |
| Base model | `LiquidAI/LFM2.5-230M-Base` |
| Maker | Kestrelyn |
| Released | 2026-10-03 |
| License | LFM Open License v1.0 (commercial use only below US$10M revenue; non-commercial above that) |
| Reported accuracy | 64.5% |
| Reported latency | 19 ms p50 / 126 ms p95 on one RTX 3060 12 GB, one request at a time |

## Get the weights

```bash
pip install systemonemodels
systemone pull kestrelyn/kestrel-decider
```

The files are served from the maker's Hugging Face repository, [`Kestrelyn/kestrel-decider-230m`](https://huggingface.co/Kestrelyn/kestrel-decider-230m), and verified against the checksums recorded here.

## Read more

- [Model card](https://huggingface.co/Kestrelyn/kestrel-decider-230m)
- [Recipe (strands-decider)](https://github.com/strands-labs/strands-decider)
- [Base model](https://huggingface.co/LiquidAI/LFM2.5-230M-Base)

---

*This page was opened by System One for Kestrelyn, who can claim the organisation and take it over at any time.*

---

From System One Models — https://systemonemodels.ai/ · every System One model: https://systemonemodels.ai/system-one-models