Jev: TypeSafe’s new AI model is not an LLM
A new type of artificial intelligence model aims to change how software uses AI. Jev, developed by TypeSafe AI, is not a traditional Large Language Model (LLM) and is not designed for conversations. Instead, the model produces structured decisions, accompanied by probabilities and confidence estimates, with the goal of making software automation faster, cheaper, and more predictable.
The idea quickly caught the attention of developers, and the emergence of Jev marks a possible shift in the development of AI-based applications. The information was presented by TechCrunch on September 18, 2026, in an article dedicated to the TypeSafe project and Diogo Almeida, a researcher who previously worked at OpenAI.
What is Jev and why is it different from an AI chatbot?
Most popular AI models are built to generate text. ChatGPT, Claude, Gemini, and other linguistic models receive information and produce sequences of text, making them extremely flexible for conversations, programming, writing, and analysis.
Jev approaches the problem differently.
The model developed by TypeSafe AI is designed for structured decisions that software can use directly. Instead of generating a textual response that must then be interpreted and validated by a program, Jev returns structured values and probabilities associated with decisions. TypeSafe refers to this category of AI models as “System One Models”.
The concept is important for automation. If an application needs to decide whether a message is spam, whether a transaction requires human verification, or whether a security incident needs to be escalated, it does not always need a paragraph generated by AI. It needs a clear decision.
Why did TypeSafe create an AI model that does not generate text?
According to Diogo Almeida, the founder of TypeSafe, the problem relates to the difference between what is optimized for humans and what is useful for computers.
Almeida previously worked at OpenAI and is associated with the development of methods such as Reinforcement Learning from Human Feedback (RLHF), used to improve the behavior of linguistic models. In his view, models have become very good at language, but natural language is not always the ideal interface for software.
TypeSafe thus aims to build an AI model primarily intended for use by other programs.
The company states that Jev uses a new architecture, a parallel sampling mechanism, and a training method called Reinforcement Learning for Calibrated Decisions (RLCD). Instead of optimizing the generation of responses preferred by humans, the goal is to achieve structured decisions and calibrated probabilities.
How can Jev reduce the problem of AI hallucinations?
One of TypeSafe’s central arguments is that Jev does not generate text. In a traditional LLM, the model can produce an incorrect statement, faulty code, or invented information. The software must interpret the response and check if it meets the expected format.
In the case of Jev, the results are defined in advance. The model can select, for example, one of several options or provide a probability for a certain decision.
TypeSafe claims that this approach eliminates type errors because the structure of the output is defined before the model produces the decision. The company also states that each result is accompanied by an estimate of confidence.
However, an important distinction is necessary: the fact that the model produces structured results does not mean that all its decisions are automatically correct. Users must set thresholds for automation and the situations in which a decision must be sent for human verification.
Is Jev faster than traditional AI models?
Speed is one of TypeSafe’s main promises. The company claims that Jev can respond in approximately 70-500 milliseconds, depending on the task, and that its approach can be tens or hundreds of times faster in certain automation scenarios. These values are assertions from TypeSafe and should be interpreted in the context of tests and workflows presented by the company, not as a universal rule for all AI applications.
The publication TechCrunch reports concrete examples provided by developers. A software engineer from Vercel stated that in a classifier for verifying order safety, switching from an OpenAI model to Jev led to results that were 5 to 18 times faster and greater accuracy in that test. Another developer compared Jev to Gemini for classifying commercial emails and reported a trade-off between accuracy, cost, and the availability of confidence scores.
These are results from specific tests, not independent benchmarks demonstrating the overall superiority of the model.
How much does Jev cost?
Low cost is another important component of TypeSafe’s strategy. On the official product page, the company indicates a price of $42 for one billion input tokens, and output tokens are presented as free. TypeSafe claims that the architecture is much more efficient for decision-making tasks than traditional linguistic models.
For applications that process very large volumes of data, the cost difference can become significant.
For example, a system that classifies millions of documents, constantly checks data streams, or monitors AI agents may need millions or billions of evaluations. In such situations, every reduction in cost per operation can change the economic feasibility of a product.
What can Jev be used for?
Jev does not aim to replace all uses of a chatbot. The use cases presented by TypeSafe are focused on automation and decisions within software.
Examples include:
- automated classification of documents and messages;
- invoice processing;
- detection and prioritization of security incidents;
- monitoring AI agents;
- detection of jailbreak attempts;
- routing requests to the appropriate AI model;
- risk assessment;
- scoring systems;
- software flows that need to choose between multiple actions;
- real-time applications where latency is critical.
TypeSafe also presents workflows for security incidents, invoice processing, customer service, and observability of AI agents.
Jev can work alongside other LLM-type models
One interesting idea is that Jev does not necessarily have to replace linguistic models. The model can be used as a verification layer for other AI agents. For example, an LLM can generate a response or execute a series of actions, and Jev can analyze the result and decide whether it should be accepted, rejected, or sent to a human operator.
This architecture can be relevant as applications use more and more autonomous AI agents. Instead of each agent being manually checked, a specialized decision model can function as an automatic filter.
What does “System One Model” mean?
The name comes from the distinction popularized by psychologist Daniel Kahneman between fast, intuitive thinking and slow, deliberative thinking.
TypeSafe uses the concept to describe models focused on quick and structured decisions, as opposed to models designed for extensive reasoning and language generation.
Jev is the first public model in this category developed by TypeSafe and is currently in early access.
What should be watched before considering Jev as an alternative to LLM?
The most important question is whether specialized models can turn promises of efficiency into consistent results in production.
TypeSafe publishes its own evaluations and explains the methodology used. The company also warns about the limitations of evaluations and claims that public benchmarks can create incentives for artificial optimization of results.
Therefore, claims such as “193.6x faster” or “444.6x cheaper” should be read as results reported by TypeSafe for certain workflows, not as a universal characteristic of Jev in any application. The company’s website presents these values in the context of its own tests.
For developers, testing on their own data and workflows remains essential.
Why is this model important for the future of artificial intelligence?
Jev illustrates a broader idea: the future of AI does not have to consist exclusively of models that converse with humans.
In certain applications, natural language can be an unnecessarily costly interface. If a program only needs the response “approved,” “rejected,” “develop,” or a distribution of probabilities, generating an entire paragraph may be an unnecessary intermediate step.
The TypeSafe model aims to move artificial intelligence closer to code and application logic: data comes in, the model evaluates the situation, and the software receives a structured decision.
This approach could become relevant in enterprise automation, cybersecurity, document processing, AI agents, and applications with strict latency requirements.
For now, Jev is a product in its early stages and available in early access. However, its emergence highlights an important direction in the AI industry: not every problem needs a chatbot, and some of the most useful AI models for software may be the ones that do not generate text at all.
Sources: typesafe.ai, techcrunch.com
