TypeSafe AI launches Jev, a digital output LLM

Jev, launched by TypeSafe AI, processes a textual "state" and responds with numbers rather than text. The model only charges for input tokens, with free output and a stated price of $0.042 per million tokens. Within days, the community has ramped up experiments, hacks, and attempts at reimplementation, amid warnings about opacity and biases.

Numbers Without Justification and Difficult-to-Isolate Biases

Jev highlights the "black box" nature of machine learning, as it only provides numerical values without any explanation for the decisions made. Concerns have been raised regarding potential biases, including a recommendation not to use Jev for job application screening and the difficulty in identifying possible hidden biases. In one test, Jev assigned the highest scores to Cupertino and the lowest to East Palo Alto for the question "Is this a good city?" applied to cities in the San Francisco Bay Area. In this context, it is deemed necessary to resort to structured evaluations and experimental protocols. The low cost of the service allows for hundreds or thousands of experimental queries for just a few cents.

What Jev Can Do: Classification, Sorting, and Reranking

The decision model framework applies to classification tasks, such as spam detection, label suggestion, prioritization, or ranking. One experiment involved retrieving 100 candidate documents using an algorithm like BM25, then asking Jev to score these 100 proposals based on their relevance to the initial query.

Text Input, Probabilistic Outputs, and Parallel Evaluation

Jev takes as input a "state," which can be a string, an array of strings, or name-value pairs describing, for example, an article or a client, accompanied by one or more questions. The model provides a numerical score for each question. Three types of questions are available: binary questions, referred to as "Noul," which the CEO of TypeSafe AI indicates is named after the Bernoulli distribution; multiple-choice questions, which generate a chosen option, a confidence level, and a probability distribution across all options; and score questions, for which the answer corresponds to a floating-point number within a predefined range. The API supports one state document and as many questions as the context window allows, processing these questions simultaneously so that batch sending does not significantly increase response time.

A Model Charged Only for Input, at $0.042 per Million Tokens

Unlike traditional LLMs that charge for both input and output tokens, often with more expensive outputs, Jev only charges for input and makes output free. The announced price for its first model is $0.042 per million tokens, which is lower than the $0.05 per million tokens for OpenAI's GPT-5 Nano. The service is also presented as fast and cost-effective.

Users Repurpose Jev to Create Novel Uses

In the days following the launch, the community has multiplied repurposed uses. Kyle Pena transformed Jev into a chat model by predicting the "next symbol" at each step, an approach likened by a Hacker News user to "Morty talking with the death crystal." Fatih Kadir Akın implemented left-pad via a bounded choice question between "0 spaces needed" and "10 spaces needed," while Andy Gayton used Jev to play the game 2048. Meanwhile, attempts at reimplementation with open-weight models have emerged: Kev, based on Qwen 3.5, offers variants of 0.8B, 4B, and 9B parameters, while a Hacker News thread references JevBench, a benchmark dedicated to these "Jev-class decision models." Jev was unveiled a week ago by TypeSafe AI as "System One model," a designation that some prefer to call "decision model."