Jet

A family of typed decision models. Give one a state and named questions; it returns choices, scores and probabilities, never free-form text.

Jet-4BLatest model, v6.2

bf16Merged weights, 8.4 GB

40.0Decision Index 0.3, #52 of 113

How it works

Each answer option maps to a single label token. Jet reads the next-token logits for those labels only, divides by a temperature fitted on held-out data, and applies softmax. One forward pass per question, no sampling. Answers always follow the requested type, but decisions can still be wrong.

Type	Criteria	Answer
`choice`	2–255 named options	Selected key, probability per key
`score`	2–10 ordered levels	Fractional score, level, probabilities
`noul`	None	Probability that the answer is yes

Fine-tuned from Qwen3.5-4B with a rank-16 LoRA, merged into full weights.

Benchmarks

Decision Index 0.3

The official leaderboard, run by its maintainers on the full suite: 110,201 requests across 42 benchmarks, published October 7, 2026. Models on the same 4B base shown; Jet-4B v6.2 scores 40.01, rank 52 of 113.

The index weights public benchmarks 20%, same skills 50% and new domains 30%, after equating each part across models. Scores within 0.9 points are tied, as Jet-4B is with its neighbors. It answered 109,958 requests; 243 exceed its 16,384-token prompt limit. Details · Leaderboard

Show as table

Model	Index	Public	Same skills	New domains
Hopper (G) 1.2	42.61	41.15	42.37	39.61
JPT-4B	41.63	42.82	40.11	39.17
Jet-4B v6.2	40.01	42.17	40.93	34.46
Kev 4B r10	39.50	37.95	36.17	41.25
InternLM Intern-Decision	38.21	38.00	39.10	34.35
Interfaze lev	36.53	39.20	37.20	31.98
Together Tev1	34.14	29.51	31.78	37.65
Jobe	33.06	32.93	34.34	30.08
Qwen3.5-4B Base (SemIf)	28.84	26.50	28.04	30.68

v6.2 by area

Decision Index skill score per area.

Strongest at tools and retrieval, weakest at knowledge and reasoning. Calibration: 66% accurate at 80% mean confidence (ECE 0.14), so confidences run high.

Examples

Real outputs recorded in the model's reference file (from Jet v5, an earlier 0.6B model, calibrated). Pick a case to see the request and the typed answer.

Die Feststellung Die Feststellung is an East German film. It was released in 1958.

choice What kind of entity is described?

Film

Film97.8%

NaturalPlace0.5%

Artist0.5%

Company0.4%

OfficeHolder0.3%

3 other options0.6%

Response JSON

{
  "type": "choice",
  "choice": "Film",
  "probabilities": {
    "OfficeHolder": 0.0027,
    "Plant": 0.0023,
    "Company": 0.0041,
    "Film": 0.9779,
    "NaturalPlace": 0.005,
    "Artist": 0.0047,
    "MeanOfTransportation": 0.002,
    "EducationalInstitution": 0.0013
  },
  "confidence": 0.9292
}

For live inference, run the model locally; see Run it.

Run it

The v6.2 release is self-contained: merged bf16 weights plus the CUDA runtime.

hf download quaedra/jet --revision v6.2.0 --local-dir jet-4b
cd jet-4b && python -m pip install -r requirements.txt
echo '{"state":"I was charged twice this month.",
  "questions":{"billing":{"type":"noul",
  "instructions":"Is this a billing issue?"}}}' | python jet.py