Meet 'Jev,' an AI That Returns Only Answers Programs Can Use Directly

TypeSafe AI unveiled a new AI model called Jev on September 15 (local time). Instead of producing prose for people to read, it returns only the decisions that software can act on directly. It is the first product in a new category the company calls System One. A program sends along the context and the shape of the answer it wants, and Jev returns a selected option, a score, or the probability that something is true. The company has opened early access and is taking developer signups through a waitlist.

An AI that answers with options instead of sentences

Jev does not generate free-form text. It reads input and replies only with predefined options or numbers. Multiple questions can be sent at once and handled in parallel. Selections and scores come with the probability that the result is correct and a confidence level. When you wire this into a service, what matters is not a human reading the AI’s explanation. It is feeding that result straight into the next step of the program.

A good illustration of the difference is llm-typesafe, a connector tool released by Simon Willison. The example shown asks whether a customer message contains a refund request and returns, as a number, the probability that it does. That corresponds to the triage step in customer support, where you sort out what kind of request came in. Willison also described Jev as a model focused on judgment.

Much faster and cheaper, the company says. Is it?

Jev leads with speed. Users are charged only for input, with no charge for output. The idea is to cut waiting time and per-call cost in software that repeats short judgments countless times. The company said it applied a new model architecture and a method for obtaining multiple results at once. It also explained that the model was trained so that its confidence levels line up with actual accuracy rates.

The company said Jev is 20 to 200 times faster and 40 to 400 times cheaper than recent small models. All of these are figures the company measured itself, under specific conditions. When weighing adoption, you are better off measuring the latency and cost on your own service’s queries than leaning on those multipliers.

Use cases already emerging

Reactions since launch have gone in two directions: bolting Jev onto existing tools, or experimenting directly with a similar judgment-style approach. llm-typesafe makes Jev usable from existing conversational AI tooling. Projects such as simple-jev and jev-align have gone up on GitHub one after another. Several curated lists of related resources have also appeared.

A Korean-focused project has also emerged. Malkuth, introduced on GeekNews on September 24, bills itself as a model specialized in multilingual and Korean classification. According to the developer’s description, it takes text, a question, and a set of options, and returns a probability for each option. Two models of different sizes were released, and the developer said they can be used together with TypeSafe’s connector approach. It is an attempt to port Jev’s judgment-style approach to Korean-language tasks.

So can you trust it in production?

Answering in a fixed format and producing the correct answer are two different things. The company’s claim that there are no formatting errors does not mean the judgments are always right. The company’s own documentation recommends breaking questions down to be narrow and specific. Tasks that require extended reasoning, or that bundle several judgments into a single conclusion, should be handled in your program code, it says.

The shift Jev points to is an approach that slices the work you hand to AI into small, unambiguous judgments. What matters to users is how quickly and accurately those judgments work in a real service. The yardstick for gauging adoption in Korea also comes down to accuracy and confidence on Korean-language queries. Whether experiments like Malkuth are usable for real work should be examined by that same standard.