TypeSafe AI released Jev last week, a model that breaks from the standard LLM format by returning numbers instead of text. The company calls it a "System One model," though decision model captures the idea more clearly.
How Jev works
You send Jev a "state" object containing text, an array of strings, or key-value pairs, along with one or more questions. The model processes the input and returns floating point numbers with confidence scores. Questions run in parallel, so fifty queries take roughly the same time as one.
Three question types are available. Noul questions (short for Bernoulli) accept a yes/no statement and return a score between 0 and 1. Choice questions pick one option from a provided list and return a probability distribution across all options. Score questions accept a numeric scale with descriptions and return a value along that range.
Pricing and performance
Jev charges only for input tokens at $0.042 per million, with output free. That undercuts OpenAI's GPT-5 Nano at $0.05 per million. The low cost makes it practical to run hundreds or thousands of evaluation prompts for a few cents.
Where decision models fit
Jev works well for classification tasks: spam detection, label suggestion, prioritization, and ranking. Search reranking is another strong use case, where you fetch a hundred candidates with a fast algorithm like BM25 and let Jev score them for relevance against the original query.
The model does not explain its reasoning. You get a number, nothing more. Standard LLMs at least let you ask for justification, even if the explanation may not be reliable. Jev removes even that thin layer of transparency.
Bias concerns
The black box nature makes bias a real worry. In one experiment, Jev rated every Bay Area city on a "good city" yes/no question. Cupertino came out on top, East Palo Alto at the bottom. The result raises questions about what patterns the model learned and how they might affect higher-stakes decisions like hiring.
Structured evaluations become even more important with a system that offers no insight into its internal reasoning. The low cost helps here too, since running extensive evals does not require a large budget.
Community experiments
The release spawned several projects within days. Kyle Pena's jevchat turns Jev into a chat model by asking it at each step which symbol comes next given the conversation so far. Ericpruitt on Hacker News compared it to "Morty speaking with the death crystal." Fatih Kadir Akin built jev-leftpad, which asks Jev how many spaces are needed and returns from zero to ten options. Andy Gayton used it to play the 2048 sliding puzzle game.
Open weight alternatives
Developers quickly began building similar models on top of open weights. Kev, built on Qwen 3.5, offers 0.8B, 4B, and 9B parameter variants. A JevBench benchmark has appeared to compare decision models against each other. The pace of activity in the first week suggests the concept has real appeal.
Decision models carve out a specific niche. They are not general-purpose assistants or creative writing tools. They are fast, cheap classification functions that happen to use neural networks under the hood. For teams processing large volumes of text and needing structured outputs, that trade-off may be exactly what they need.