# Jared Palmer releases Kev-9B, an Apache-2.0 LoRA decision model on Qwen3.5-9B

> An independent adapter that implements TypeSafe's System One API. The card reports 0.852 on a locked out-of-domain test.

Published 2026-09-20

Canonical: https://shipped.anth.us/decisions/jaredpalmer-kev-9b/

Jared Palmer published Kev-9B on Hugging Face on 20 September 2026, by repository creation date. The card describes a decision model that reads one document and a set of typed questions and returns calibrated probabilities over the supplied options in a single forward pass, without generating text. It implements TypeSafe's System One API.

It is a LoRA adapter (rank 16, about 45.4M parameters) plus a pointer head on Qwen/Qwen3.5-9B-Base. The adapter and head are Apache-2.0, as is the base. It fits a 24 GB-class GPU, with about 22 GB resident on CUDA. Code is on GitHub.

The card reports accuracy of 0.698 on a 14-dataset public breadth test, 0.852 on a locked out-of-domain transfer test (Brier 0.199), 0.834 on a hard test, 0.791 on a devtools test and 0.900 on CFPB complaints. On MMLU-Pro with ten options it reports 0.590. It also says v2 is not better than v1 on new work, and that the large gains are on families it trained on.

Limits listed: English only, validated context 8,192 tokens, weakest on date arithmetic, calibration may not transfer. It is not for consequential automated decisions about people without human review.

The card says what did not improve. That is the part to read.

## Sources
- https://huggingface.co/jaredpalmer/kev-9b
- https://github.com/jaredpalmer/kev

## Editor's first pass (AI-assisted)
- Real 4: Adapter, head and code are public; fits one 24 GB-class GPU.
- Useful 3.5: Document classification and routing; card reports 0.900 on CFPB complaints.
- Notable 3: Other releases (Decision4Rec Kev) build on it.
- Novel 3: An independent take on an existing API.