Jared Palmer's Kev-4B Skips Text Generation to Score Decisions Directly
A 4B open-weight decision model that returns calibrated probabilities to typed questions in one forward pass, priced at $0.042 per million input tokens.
Takeaways
- Kev-4B is live on OpenRouter at $0.042/M input tokens, $0 output, 8K context.
- LoRA adapter plus pointer head on Qwen3.5-4B-Base, Apache-2.0, weights on Hugging Face.
- Returns typed answers with calibrated probabilities in one forward pass, no text generation.
- Scores 0.837 out-of-domain accuracy on locked test with built-in temperature calibration (T=2.14).
- Weakest at knowledge tasks (MMLU 0.70 vs Jev 0.90) and raw date arithmetic without preprocessing.
- Drop-in replacement for Jev: change one field, keep the same
/v1/systemoneclient code.
Jared Palmer has released Kev-4B on OpenRouter. The 4B open-weight model accepts a document and one or more typed questions, then returns a calibrated probability distribution over each question’s answer options. The model bypasses token-by-token text generation.
Developers already using TypeSafe’s Jev can keep the /v1/systemone contract and change the model field to "jaredpalmer/kev-4b". OpenRouter charges $0.042 per million input tokens and $0 per million output tokens, with an 8K context window.
Why classification removes decoding overhead
Routing, risk scoring, document classification, tool gating, and similar systems usually select from known answers. A generative model handles those jobs by decoding text, producing JSON, and relying on downstream validation. Kev maps the same inputs directly to answer probabilities, removing generation latency and malformed-output handling.
Kev adds a rank-16 LoRA adapter and a pointer head to Qwen3.5-4B-Base. For each question, the pointer head compares the hidden state of every answer option with the question’s hidden state. A softmax converts those scores into probabilities suitable for thresholds, ranking, or abstention.
Strong on science, weaker on knowledge
On the locked out-of-domain test, Kev-4B reports 0.837 accuracy and a Brier score of 0.255. Lower Brier scores indicate better probabilistic accuracy. Its in-distribution accuracy is 0.872.