Decision models · Source checked September 27, 2026
Jev vs Kev
Both answer bounded questions with structured probabilities. Jev is a managed API; Kev is an open model family you can train and run yourself.
Hosted API
Jev
Send state and questions to Jev and receive answers without operating a model server.
Open weights
Kev
Run a Qwen-based decision model locally or on your own endpoint, and fine-tune it for your labels.
Side by side
The practical differences
Kev follows the System One request format, including noul, choice, and score questions. The larger choice is who runs the model and how much you need to customize it.
Deployment
Managed TypeSafe API; no model infrastructure to operate.
Local server, your GPU, or a self-deployed cloud endpoint.
License
Proprietary hosted model.
Apache-2.0 project and base models; training datasets have separate licenses.
Model choice
Hosted Jev model selected through the API.
Published 0.8B, 4B, 9B, and 27B checkpoints with different hardware needs.
Interface
System One API with noul, choice, and score questions.
Compatible /v1/systemone format; TypeSafe's Python SDK can point to a Kev server.
Customization
Define your questions and criteria in each request.
Define questions and fine-tune released checkpoints on your labeled examples.
Probabilities
Returns probabilities for application decision thresholds.
Checkpoints ship with a fitted temperature; validate thresholds on your own data.
Published evaluation
Read the accuracy numbers carefully
Kev's README reports new-source accuracy on its development set. Higher is better; these are the figures the project publishes for Jev and Kev-27B.
Jev
85.7%
New-source accuracy, development set. No Jev test-set figure is published there.
Kev-27B
84.8%
New-source accuracy, development set. The README also reports 89.6% on Kev's test set.
These figures come from Kev's evaluation, not an independently controlled architecture comparison. Jev's training overlap is unknown, and only Kev has a reported test-set score. Measure both on your own labeled decisions before setting thresholds. Kev README and model table
Choosing a fit
Start with your operating constraints
Choose Jev for a managed API
You want to integrate structured decisions without provisioning GPUs, serving checkpoints, or managing model updates.
Choose Kev for control
You need open weights, local deployment, or a model you can fine-tune with your own labeled decisions, and can operate the serving stack.
Kev is an independent project by Jared Palmer. Jev AI is not affiliated with or endorsed by Kev or TypeSafe. Repository details and figures were checked on September 27, 2026.