Kev uses Qwen3 to answer typed questions in one pass

Jared Palmer has expanded Kev, an open-source implementation of TypeSafe’s Jev decision-model pattern, into three Qwen3-based sizes: 0.6B, 4B, and 8B. Each model accepts a document and a set of typed questions, then returns probability distributions in one forward pass without token-by-token decoding.

The design targets ticket routing, document classification, risk scoring, and other workflows with fixed output schemas. Kev treats those jobs as classification tasks, avoiding the latency and parsing overhead of prompting a generative model to produce JSON.

One prefill, isolated questions

Kev adds a LoRA adapter and a small readout head to a Qwen3 backbone. At deployment, the corresponding Qwen3 model runs with Kev’s adapter and head.

The model packs the state, such as a document or support ticket, and every question into one token sequence. A block-causal attention mask lets each question attend to the shared state while blocking attention among sibling questions. The model can therefore score all questions in parallel without one answer influencing another.

A pointer head compares each option’s hidden state with a special <decide> token. Softmax converts those scores into a probability distribution, and cross-entropy training aligns the output with labeled examples.

The interface supports three question schemas:

• noul: a yes-or-no decision
• choice: a selection among 2 to 255 options
• score: an ordered scale, such as a 1-to-5 rating

Each question branch restarts its position IDs after the shared state, so additional questions incur branch-token and readout costs without re-encoding the document. When several requests reuse the same state, a key-value cache can retain that shared computation; Palmer reports speedups of 2 to 2.5 times.