Jev in plain English
A normal language model is optimized to continue a conversation or generate text. Jev is optimized to answer many constrained questions about the same input. You give it a state, define the decisions you need, and receive structured answers your software can act on.
A useful shorthand is: LLMs write; Jev judges. That shorthand is incomplete, but it captures why Jev belongs inside software workflows rather than in a consumer chat box.
The three output primitives
- Choice selects one option from criteria you define, such as billing, technical, or sales.
- Score places the input on an ordered rubric, such as low, medium, or high risk.
- Noul estimates whether a property is present and returns a probability-like value.
Why it is fast and inexpensive
Jev gives up free-form string generation. It can evaluate its outputs in parallel instead of generating a response token by token. TypeSafe currently lists end-to-end latency of roughly 70–500 ms for its examples and prices input at $0.042 per million tokens; output tokens are free.
Those numbers come from TypeSafe and depend on the task, request size, location, and service conditions. Treat them as product claims to validate against your own workload.
What Jev cannot replace
- It does not write articles, emails, code, or explanations.
- It currently accepts text only—no image, audio, or video input.
- It works best when the possible decisions are known before the request.
- English is its strongest language; other languages require workload-specific testing.
- Complex planning and open-ended creation still belong to an LLM or another system.
Where Jev fits
Jev works well as a fast semantic decision layer between ordinary code and a larger model. Common patterns include routing support tickets, ranking candidates, filtering content, choosing agent tools, retrieving context, and checking another model’s output.
Continue reading
Next: How to use Jev
From the official console to your first API request.