A small decision model inspired by Jev. You send a state and some named, typed questions, and you get typed answers and probabilities back. Jet doesn't generate text, so there's nothing to parse and it can't return an answer outside the type you asked for.
Explore the demos · Read the docs
Current release: Jet v6.2.0 · full merged Qwen3.5-4B model. The charts below show the earlier v6.1 evaluation across 25 benchmarks, including two retrieval samples. Jet has no official overall index score.
Chart scores measure the v6.1 step-2,000 adapter before its BF16 merge; they are not v6.2 scores. Matching metrics and case counts do not verify identical source cases; this is not an official ranking.
914 local holdout examples, evaluated on the full merged model. These are separate from the benchmark charts above; financial sentiment uses SEntFiN, not FinEntity. Focus holdouts are new local splits; retention is reused. Small changes do not establish statistical significance.
| Holdout metric | v6.1 | v6.2 |
|---|---|---|
| Banking accuracy | 74.68 | 74.68 |
| Financial sentiment F1 | 71.16 | 71.21 |
| Sarcasm F1 | 46.81 | 50.00 |
| Retention accuracy | 93.40 | 93.40 |
Current Jet model card · Full-model evaluation
| type | criteria | answer |
|---|---|---|
choice | {key: description}, 2–255 | a key, probabilities per key, confidence |
score | [level, …], 2–10, low→high | fractional score, argmax level, probabilities |
noul | none | probability that the answer is yes |
POST /v1/decide
{
"state": "I was charged twice this month and nobody answers my emails.",
"questions": {
"topic": {"type": "choice", "instructions": "What is the primary issue?",
"criteria": {"billing": "billing or payment problem",
"bug": "the product is broken"}},
"escalate": {"type": "noul", "instructions": "Escalate to a human immediately?"}
}
}{
"answers": {
"topic": {"type": "choice", "choice": "billing",
"probabilities": {"billing": 0.97, "bug": 0.03}, "confidence": 0.8},
"escalate": {"type": "noul", "probability": 0.81, "confidence": 0.3}
}
}Each decision is one forward pass with no sampling. Every option maps to a single label token
(A…Z for choices, 0–9 for score levels,
yes/no), and the answer is the model's next-token distribution over only
those tokens. The release uses inherited temperature scaling for its probabilities. Each complete question is processed separately, with a 16,384-token limit and no input truncation.
The base model is Qwen3.5-4B. The trained LoRA updates are merged into the full Jet weights, so no separate adapter or base-model download is required.
Jet is trained on public dataset training splits and deterministic examples for tasks such as arithmetic, Boolean rules and code behavior. Dataset sources, revisions, training recipes and evaluation audits are documented in the training records.
The source code includes data builders, training scripts, evaluation tools and the inference server. Download the model weights to run Jet locally, or explore the website source.