the short answer
Jared Palmer’s Kev-9B pairs a Qwen3.5-9B-Base model with an Apache 2.0 adapter and pointer head. The project documents a System One-compatible local server and a 24 GB-class GPU setup. Jev is TypeSafe’s separate hosted model. Kev’s calibration setting and published scores belong to its own checkpoint. Run contract tests and fit new action thresholds before swapping it into a Jev workflow.
- Publisher
- Jared Palmer; independent of TypeSafe AI
- Checkpoint
- Kev-9B v2, described in the September 30 model-card update
- Architecture
- Qwen3.5-9B-Base plus LoRA adapter and pointer head
- Serving
- Project documents a
/v1/systemone-compatible local server - Scope
- English text decisions; not a chat or generation model
The Compatible Server and the Checkpoint Behind It
Kev’s card shows Choice, Score and Noul questions returning distributions through a local /v1/systemone server. The author also documents separate weights, training sets and a calibration temperature, and says Kev did not train on Jev outputs. The shared route shape helps with client experiments. It does not make the model Jev.
Version 2 uses Qwen3.5-9B-Base, a LoRA adapter and a pointer head. The card validates English and 8,192 tokens, even though the server accepts a larger maximum. Keep long input cases in the test set rather than reading the maximum as a quality promise.
A Local Policy-Check Fixture
Start this policy check in observation mode. A model may miss a path, amount or permission condition that code can check exactly, and missing authorization deserves its own test cases. The Jev policy guide explains when a judgment can feed review or enforcement.
| Field | Test value |
|---|---|
| State | Proposed tool action, user authorization and the relevant policy text |
| Noul | Does the proposed action violate this named policy condition? |
| Code checks | Path, permission, amount and exact command invariants |
| Outcome | Observe or review until false blocks and false allows are measured |
| Audit record | Checkpoint hash, provider, question version, distribution and later human outcome |
Compatible Endpoint, Separate Thresholds
Kev’s card reports several datasets. Keep their metrics beside the dataset and version that produced them. A Jev number from a different suite cannot complete the comparison. Pin both models before running your own cases.
| May carry over | Must be re-earned |
|---|---|
| Question names and answer definitions | Response-field and error conformance |
| Labeled cases and policy rubric | Accuracy, Brier score and action bands |
| Application allow/review/block contract | Timeout path, hardware capacity and rollback |
How to Evaluate Local Kev Against Hosted Jev
Use the open-model guide to check the checkpoint and license. Save the paired responses, labels and failures using the benchmark method.
- Pin Kev model revision, base revision, serving code and calibration setting.
- Pin Jev model, provider, state projection and question version.
- Run a wire-contract suite for every primitive and failure response.
- Evaluate blind held-out labels with separate thresholds and risk-coverage curves.
- Include GPU, queueing, retries, review work and maintenance in total cost.
FAQ
Is Kev-9B an official Jev release?
No. Jared Palmer publishes Kev-9B independently. Its TypeSafe-compatible endpoint does not make it a TypeSafe model.
Can Kev-9B run locally?
The project documents local serving with its adapter, head and base model on a 24 GB-class GPU. Verify hardware, license and exact revision for your deployment.
Do Jev thresholds transfer to Kev?
Fit new thresholds for Kev on your labeled cases. The shared zero-to-one output range does not tell you how often each model is right at 0.8.
Sources
Checked against the sources below on October 2, 2026. Model versions, prices and limits change.
- Jared Palmer: Kev-9B model card
- Jared Palmer: Kev source and server
- TypeSafe AI: Jev primitives
- TypeSafe AI: Jev models
Change note: Added Kev-9B v2 provenance, local-serving facts and substitution tests.