In one line: Jev does not write a long answer. It makes a bounded decision inside an answer space you define.
How it works
A chat LLM usually generates text token by token and leaves your code to parse it. Jev separates state, questions and answer spaces so the result can be handled directly by software.
state + typed questions
↓
choice / score / noul + probabilities
↓
action / confirm / escalateThree primitives
Choice
Pick from options
Returns a candidate and a probability for each option.
Score
Rate against a rubric
Returns a position and the distribution across levels.
Noul
Ask yes or no
Returns the probability that the statement is true.
Good fits
- Support routing and agent tool selection;
- Moderation, risk triage and human escalation;
- Batch classification, ranking and quality checks.
Bad fits
Long-form writing, code generation, summaries, exact arithmetic, date calculations and open-ended explanations belong elsewhere.
Independent-site note: Community projects may reproduce an interface, inference path or training idea. They do not become official Jev weights.
Next step
Start with the API quickstart or browse the open model directory.