AI has become remarkably good at generating text. TypeSafe wants to change what happens next.
TypeSafe AI has introduced Jev, the first model in a new category the company calls System One Models. Instead of generating text, code, or long-form reasoning, Jev is designed to make fast, structured decisions that software can use directly.
The release marks a different direction from the traditional language-model approach. Large language models are built primarily around generating sequences of text. Jev is built around a simpler idea: give software the information it needs to evaluate, ask a defined question, and return a structured decision.
Meet Jev
Jev was released by TypeSafe on September 15, 2026, following two years of development in stealth. TypeSafe describes it as its first public System One Model, a new class of models designed specifically for automation.
The distinction is fundamental.
A conventional AI model might receive a request and respond with a paragraph explaining what it thinks. That response then has to be interpreted, parsed, validated, and converted into something a software system can safely act on.
Jev skips that process.
Its input is a state and a set of typed questions. Its output is a structured result that software can consume directly.
In TypeSafe's terminology, the idea is simple: unstructured state in, typed probabilistic decisions out.
Decisions instead of generated text
Jev is built around three main question types: Choice, Score, and Noul.
Choice selects between predefined options. Score evaluates something against a defined scale. Noul answers whether a statement is true using a probability between zero and one.
Each decision can also include probabilities and confidence information, allowing software to account for uncertainty instead of treating every AI response as equally reliable.
That changes the role of the model.
Jev does not need to explain its reasoning in a paragraph. It does not need to generate JSON from a text response and hope that the structure is correct. The possible output is defined ahead of time, allowing the surrounding software to work with the result directly.
Built for software, not just people
TypeSafe's argument is that traditional language models were created around a human interface.
Humans can read a generated answer, understand it, and decide whether it makes sense. Software does not work that way. A program needs predictable types, explicit outcomes, and clearly defined behavior.
Jev is designed around that requirement from the beginning.
TypeSafe says Jev produces type-safe structured values, meaning the model cannot return a value outside the structure defined by the application. Every answer also includes probability information, giving software a way to distinguish between confident and uncertain decisions.
The company positions this as a new primitive for AI-powered software: something closer to an intelligent function call than a traditional chatbot response.
Speed is a major part of the design
Jev is also designed differently at the sampling level.
Traditional language models generate text sequentially, producing one token after another. Jev instead evaluates its outputs in parallel, which TypeSafe says allows the model to operate far more efficiently for this type of structured task.
TypeSafe currently reports end-to-end response times in the range of 70 to 500 milliseconds for its service.
That is important because AI inside software often has very different latency requirements from AI used as a conversational interface. A response that takes several seconds may be acceptable for a chatbot, but becomes much more difficult to use when an application needs to make frequent decisions in real time.
A different approach to AI economics
The same design also affects the cost of inference.
TypeSafe lists Jev at $0.042 per million input tokens, or $42 per billion input tokens, with output currently listed as free.
The company argues that removing text generation dramatically reduces the amount of computation required for these structured decision tasks.
That creates an interesting possibility for applications that need to make large numbers of small AI decisions rather than generate large amounts of content.
Confidence is part of the output
One of the more interesting aspects of Jev is that uncertainty is treated as part of the result rather than something the developer has to infer from generated language.
Choice and Score responses include confidence information alongside their decisions. TypeSafe describes these probabilities as calibrated, meaning the intention is for higher confidence to correspond to a higher likelihood that the decision is correct.
For software, that creates another option beyond simply accepting or rejecting a model response.
An application can decide how it wants to behave at different confidence levels, including acting automatically when confidence is high and using another path when the model is uncertain.
Where Jev is designed to fit
TypeSafe is positioning Jev around tasks such as classification, routing, scoring, verification, extraction, guardrails, and other situations where software needs an AI judgment rather than a generated explanation.
That includes AI-powered workflows where multiple small decisions are combined inside normal application logic.
The model is also designed to answer multiple questions against the same state in parallel. TypeSafe says each question is evaluated independently, allowing developers to decompose a complex decision into smaller, more controlled questions rather than asking one model call to solve everything at once.
Not another chatbot
Jev is therefore not trying to compete with language models on the same ground.
It does not generate essays. It does not write code. It does not attempt to replace a conversational assistant.
Its purpose is much narrower: making AI decisions that software can consume directly.
That narrower scope is exactly what makes the release interesting.
Over the past few years, AI development has largely focused on making models better at understanding and generating language. TypeSafe is exploring a different layer of the stack, where intelligence becomes a structured component inside ordinary software instead of the software simply talking to an AI through a text interface.
The beginning of System One
Jev is only the first public model in TypeSafe's new System One approach.
The company says it developed a new model architecture, a parallel sampling system, and a training method called Reinforcement Learning for Calibrated Decisions, or RLCD, specifically for this type of workload.
TypeSafe is releasing Jev initially through early access and is encouraging developers to experiment with the model across different kinds of automation and decision-making tasks.
The larger idea is still being defined.
But Jev represents an interesting shift in how AI models can be designed: not every intelligent system needs to produce more words. Some may be more useful when they simply make the right decision, return it in a form software understands, and do it fast enough to become part of the application itself.
Jev is TypeSafe's first step toward that vision, and System One may become one of the more interesting new directions in AI infrastructure to watch.