200× faster and 400× cheaper than LLMs
Most people think of artificial intelligence as a chatbot that writes paragraphs. Jev, a model released this month by TypeSafe AI, works differently. It does not generate text. Instead, it answers precise, structured questions about any piece of information you give it — and does so in a fraction of a second.
You provide context, called the “state” (an email, a support ticket, a webpage description, a market snapshot). Then you ask one or more typed questions: Is this true or false? Which of these options fits best? How does this rate on a scale? Jev returns probabilities for each answer. Your own software decides what to do next.
The approach is useful wherever systems need many small, reliable judgments. Support teams can route tickets automatically. Software agents can choose their next action without long delays. Moderators can score content quickly and cheaply. In each case the model runs in parallel, so asking several questions costs roughly the same as asking one.
Here is a simplified example of normal LLMs vs Jev at classifying:
🧓Consider a short customer message – “I was charged twice this month. Please refund one.”
The system asks three questions at once:
• Which team should handle it — billing, technical, or other?
• How frustrated is the customer?
• Does the message signal urgency?
A) With ChatGPT you might write a prompt like:
Classify this support ticket. Reply only in JSON with the fields: team (billing/technical/other), frustration (1-3), urgent (true/false).
The model generates text that you then parse. Latency is higher, cost is higher, and the “confidence” is not a calibrated probability.
B) With Jev the same task looks like this:
State: the customer message
Questions:
– Which team should handle it — billing, technical, or other?
– How frustrated is the customer (Calm / Mildly annoyed / Very angry)?
– Is this urgent?Jev returns probabilities for every option in one fast call. If confidence is high, the ticket is routed automatically. If not, it goes to a human.
Jev model is made for “decision-making” instead of text generation, so it generates probabilities for each option in parallel instead of giving you a response token by token.
The difference is practical. ChatGPT is flexible and conversational. Jev is built for the narrow, high-volume decisions that software systems make thousands of times a day.
🔐Jev itself remains closed. Its weights and training details have not been released. Access comes through TypeSafe’s API or resellers such as OpenRouter and Vercel.
Open-source projects already imitate the same style of interaction. The most widely used is SemIf (formerly called OpenJev), which runs on ordinary open models and even works in a browser. Others, including jevlike and mini-jev, train or adapt smaller models for the same task. None matches Jev exactly, but they let developers experiment without relying on a commercial service.
Also notable: nimble, laya-mlx, Try online: askjev.net
The result is a tool that sits between traditional classifiers and large language models: faster and more structured than the latter, more flexible than the former.