September 21, 2026
A 9B model handles document questions: Kev scores 0.852 on unseen data
Jared Palmer released the Kev family under Apache-2.0: 0,8B, 4B, and 9B answer typed questions about a document in a single pass.
Kev responds not with text, but with a probability: one model pass for a batch of questions about a document.
Jared Palmer released three models based on Qwen3.5: Kev-0.8B, Kev-4B, and Kev-9B. Each is built as a rank-16 LoRA adapter with a pointer head on top of a frozen base. Both the code and adapters are licensed under Apache-2.0.
The /v1/systemone endpoint accepts a document and a set of typed questions: yes/no, multiple choice, rating. Questions are isolated from each other, while /v1/systemone/permute shows how option order affects the answer.
What this replaces. Practitioners in a Hacker News thread name their use cases: classifying bash calls as safe or unsafe, a code-convention verifier for an agent, a spam filter instead of a 27B model. Previously, such checks required a large model. Now 9 GB of video memory in bf16 is enough for Kev-4B and 19 GB for Kev-9B; both fit in a Mac with 32 GB.
The gap on unfamiliar data. On a closed test from outside domains, Kev-9B holds 0,852, Kev-4B 0,837, while Kev-0.8B drops to 0,643. Nearly 21 points separate the smallest and largest models. The training-set composition is disclosed: 10 000 public records, 896 policy minimal-pair records, and 1680 records from 60 generated rule structures.
Installation takes three commands. Python 3.12+ and UV are required, followed by cloning the repository, uv sync --extra serve, and starting the server on port 8009. Testing uses a POST request to localhost:8009/v1/systemone, with a web playground alongside it.
Fine-tuning on your own data is inexpensive: Kev-4B trained for 56 minutes on one H100, Kev-9B for 91 minutes with a peak of 39,5 GB. The README says fine-tuning on your own data usually delivers better results than prompt engineering.
On Apple Silicon, the new generation is slower than the old one: five Kev-4B questions take 0,78 s versus 0,17 s for the Qwen3 version. Kev-9B answers in around 2 s on M5 because there are no MPS kernels for Gated DeltaNet layers.
The author has kept the previous Qwen3 generation of Kev available: on Apple Silicon, it is still faster than the new one.
Source
