/ 01
Short Answer
Jev is TypeSafe AI’s flagship model and the first System One model. You send it a state and a set of typed questions. It returns structured answers — a choice, a score or a yes/no probability — that code can use directly. It does not generate prose.
/ 02
Why “System One”?
LLMs are designed to write text for people. When software needs a judgment, asking for prose and parsing it back into a decision adds a fragile step. System One models target narrow, typed decisions that can be composed in code. This is how TypeSafe introduces Jev.
/ 03
How Jev Differs From an LLM
| LLM | Jev | |
|---|---|---|
| Output | Natural-language text | Typed JSON decisions |
| Code receives | Text to interpret or parse | Choice, score and probabilities |
| Uncertainty | May need separate measurement | Choice and Score confidence; Noul probability |
| Answer space | May write outside requested set | Restricted to defined options or levels |
| Speed and cost | Depends on model and task | TypeSafe reports gains for selected System One tasks |
Performance figures are vendor-reported and task-specific, not a universal independent benchmark. Read the use-case map.
/ 04
The Three Question Types
| Type | You provide | Jev returns | Use for |
|---|---|---|---|
| Noul | Yes/no question, optional true/false criteria | One probability, 0–1 | Checks and guardrails |
| Choice | Named options and descriptions, up to 255 | Selected option, distribution, confidence | Classification and routing |
| Score | 2–10 ordered level descriptions | Weighted score, legend, distribution, confidence | Ratings and severity |
Explore noul questions and the JSON response schema.
/ 05
How to Ask Jev a Good Question
Ask for one immediate judgment per question. “Does this message convey urgency?” is narrow. “Analyze this message and decide the entire response strategy” hides several judgments. For a pitch, ask separately about market size, feasibility and differentiation, then combine those answers with weights in your own code.
/ 06
What Jev Is Not Good At
TypeSafe’s Jev 1.13 jaggedness notes describe literal reading, counting and arithmetic, numeric representations, date comparison, indirection, irrelevant state, adversarial content, contradictory criteria, structural invariants and text generation. Read our Jev limitations guide for workarounds.
/ 07
Model Versions and Pricing
Use jev-latest for the stable alias or pin a version when calibrated thresholds matter. The response reports the actual version. As reviewed, input costs $0.042 per million tokens and output tokens are free. Verify the official model page for changes.
/ 08
What You Can Build With It
Common patterns include classification and routing, checks and guardrails, rubric scoring, large-scale labeling and verification of another model’s output. Keep control flow and arithmetic in code while Jev handles a bounded semantic judgment.
/ 09
Who Makes Jev?
Jev is made by TypeSafe AI. This site is an independent learning and experimentation tool.
/ 10
Try Jev in the Playground
Explore all three types in the Jev playground. Results clearly indicate whether they are a demo preview or live run.
/ FAQ
Frequently asked questions
What does Jev stand for?
TypeSafe’s public documentation identifies Jev as its flagship model but does not define the name as an acronym.
Is Jev a language model?
It processes text but is designed to return typed decisions rather than prose.
What is a System One model?
A model aimed at fast, structured judgments software can use directly.
What are Jev’s three question types?
Choice, Score and Noul.
Is Jev free to use?
The public API price is usage-based; this site offers free demo previews.
How fast is Jev?
It is built for low-latency decisions, but actual latency depends on workload and network. Benchmark your use case.
Does Jev make things up?
Its typed answer space prevents free-form invented labels, but a judgment can still be wrong. Validate it.