TypeSafe AI has launched Jev, an artificial intelligence model built on the Transformer architecture, but it does not belong to the large language model category and does not produce text for users. Instead, the model returns probabilities for what the company calls “calibrated decisions,” after the developer has predefined the required output types.
The company is led by Diogo Almeida, a former OpenAI researcher who helped build ChatGPT and contributed to the development of reinforcement learning from human feedback (RLHF). Almeida says AI models have focused heavily on improving their handling of human language, while software automation requires outputs that are more disciplined and directly usable by computers.
Higher Speed and Lower Cost
TypeSafe AI says that avoiding language generation makes Jev fast and inexpensive, while also reducing the likelihood of hallucinations because the user defines the possible outcomes in advance. The company does not charge for output tokens, while input tokens are measured in billions rather than millions, according to the source material.
The model quickly attracted developers’ attention, to the point that its application programming interface temporarily stopped serving users because of high demand. In a test conducted by Vercel to classify commands for security review, replacing an OpenAI model with Jev produced results between five and 18 times faster, with higher accuracy, according to software engineer Pranit Sharma.
In another test involving the classification of commercial email messages, Nikhil Mudholkar, the chief technology officer at Bryo AI, found that Gemini was slightly more accurate but cost between 10 and 20 times more. Mudholkar considered the confidence scores returned by Jev more important for automation because they provide a probability that can be used to decide whether an action should be carried out or ignored.
What Changes in Practice?
These use cases show that Jev is not necessarily intended to replace language models for every task. It could handle classification, route requests to the appropriate model, monitor the effects of AI agents, or detect attempts to bypass constraints. In these scenarios, its speed and low cost could make real-time monitoring more practical.
However, relying on probabilities does not eliminate the need for human judgment or operating rules. Armin Ronacher, the chief technology officer at Earendil, noted that users must determine how to handle a score such as 50% or 95%, meaning they must establish clear thresholds before converting outputs into automated actions. As a result, some responsibility for handling uncertainty shifts from the model to the application’s design.
Undisclosed Architecture and Expected Competition
TypeSafe AI has not disclosed details of Jev’s architecture, while outside observers suspect that it may be built on top of an open-weight language model, something the company did not confirm in the material. TypeSafe AI describes the model as a “System One model” focused on intuition and selecting the appropriate task, and says it was trained exclusively on synthetic data using a technique it calls “reinforcement learning from calibrated decisions.”
According to certi.news’s analysis, Jev’s importance lies not only in being a new model, but also in presenting AI as a specialized decision layer within software rather than as a general-purpose conversational interface. If the reported test results hold up, this approach could suit repetitive tasks with defined outputs. However, the published comparisons are limited and are not sufficient on their own to assess performance across different fields or workloads. The ambiguity surrounding the architecture and data-generation method also leaves open questions about verifiability and scalability.
TypeSafe AI plans to build additional versions of the model in different modes, and Ronacher expects competitors to emerge once the practical value of this type of model becomes clearer.