Jev and System One Models: The New LLM Paradigm Called "Decision Models"

TypeSafe AI has recently introduced Jev, the first example of what they call a new class of models: "System One models." The naming has already sparked debate in the community — including this article's own perspective — with many agreeing that Maggie Appleton's term "decision models" is a more fitting description.

Jev's core innovation lies in fundamentally changing how LLMs produce output: it still accepts text as input, but instead of returning text, it outputs floating-point values and confidence scores corresponding to classifications, yes/no judgments, and ratings. TypeSafe's official definition is elegantly concise: "Think of Jev as a function call to frontier intelligence: input unstructured state, output typed probabilistic decisions."

From Generation to Decision: A Shift in LLM Form

Traditional LLMs are "text in, text out." Jev is "text in, numbers out." This shift may seem simple, but it redefines the model's applicable use cases. It compresses the intelligence of a large model into a well-defined "decision interface," allowing developers to obtain probabilistic judgments the same way they'd call a function.

Jev supports three types of questions:

• Noul questions (yes/no): Confirmed by the CEO on Hacker News, "Noul" is shorthand for Bernoulli distribution. You present a statement, and the model returns a float between 0 and 1 representing its confidence that the statement is true.
• Choice questions (multiple choice): The model selects from given options, returning confidence scores along with a probability distribution covering all options.
• Score questions (rating): You provide a series of numerically described levels, and the model returns a float score within that range.

In practice, you assemble a "state" object — which can be a string, an array of strings, or a set of key-value pairs describing an article, a customer, or any record — then send it to the API along with one or more questions to receive responses. Notably, multiple questions are evaluated in parallel, so sending many questions takes roughly the same time as sending just one.

The Bernoulli distribution is one of the most fundamental discrete probability distributions in probability theory, describing a single random trial with only two outcomes (success/failure, true/false), with the core parameter being the "probability of success" p (a float between 0 and 1). Noul questions essentially ask the model to estimate the Bernoulli probability that a given statement is true — a concept directly in line with how traditional classifiers output softmax probabilities, except that this capability is exposed directly as an API interface rather than buried inside the model. This design allows Jev's outputs to be used directly in downstream operations like probability-threshold filtering and weighted ranking, without needing to parse natural language output a second time.

Exceptional Speed and Cost Advantages

One of Jev's most compelling features is that it's fast and cheap. Conventional LLMs are billed per input and output token, with output typically costing significantly more. Jev only charges for input — output is completely free. Its first model's input price is $0.042 per million tokens, even lower than OpenAI's GPT-5 Nano at $0.05 per million.

This pricing structure is significant for large-scale classification and scoring tasks. When you need to make judgments on tens of thousands of records, costs become almost negligible. It also directly affects practical utility — a large number of experimental calls costs just a few cents.

Use Cases for Decision Models

The "decision model" framework helps clarify where Jev excels. Any scenario expressible as a classification task is a candidate — spam detection, automatic tagging, priority ranking, and so on.

One particularly inspiring application is search reranking: first use a cheap algorithm like BM25 to retrieve 100 potentially matching candidates, then have Jev score those 100 candidates for relevance against the original query. This combination of "cheap retrieval + intelligent reranking" plays directly to Jev's speed and cost advantages.

BM25 is a classic sparse retrieval algorithm based on TF-IDF concepts, widely used in full-text search engines. It's extremely fast and requires no training, but cannot understand semantics — "Apple iPhone" and "iPhone" are completely unrelated to it. Search reranking exists precisely to compensate for this shortcoming: use BM25 or similar cheap algorithms to do coarse candidate filtering first, then use a stronger semantic model for precise reranking. Traditional rerankers (such as Cross-Encoders) need to encode both query and document simultaneously, making them computationally expensive. Jev's extremely low per-token price and parallel evaluation capability dramatically improve the economics of this pipeline, making it especially suited for online search scenarios that need to score large numbers of candidates in real time.

The Black Box Returns: A Concern Worth Taking Seriously

One unsettling aspect of Jev is that it represents a further retreat toward black-box machine learning systems. LLMs are already black boxes — you can ask them to explain their reasoning, but there's no guarantee the explanation is accurate or useful. Jev doesn't even offer that: no matter how much text you input, you get back a single float. If Jev classifies a piece of content as spam, what signals triggered that judgment? You have no way of knowing.

This means bias concerns must be front and center. The author explicitly raised this worry: hopefully no one uses Jev to rank job candidates — that float could be hiding deep-seated, hard-to-detect biases embedded in the model, and trying to unpack those biases through experimentation would be extremely difficult.

The author ran one experiment: asking Jev for every city in the San Francisco Bay Area whether it's "a good city," with Cupertino ranking highest and East Palo Alto at the bottom — results that are themselves thought-provoking.

For this reason, evaluations (evals) and structured experiments are even more critical for Jev projects than for standard LLM projects. Fortunately, Jev is cheap enough that running hundreds or thousands of experimental prompts costs only a few cents.

The interpretability problem with black-box models is particularly acute in high-stakes decision scenarios. Existing interpretability methods (such as LIME, SHAP) typically rely on perturbation experiments on model outputs to attribute feature importance, but these methods themselves have instability issues and require a large number of repeated inferences. When a model's output is a single float, attribution analysis becomes even harder — you can't ask "why" to get a natural language explanation, nor can you use techniques like Chain-of-Thought to have the model "narrate" its reasoning process. This makes using Jev in regulated contexts like recruitment screening, credit scoring, and content moderation subject to higher compliance risks and auditing challenges. Regulatory frameworks such as the EU AI Act have explicit transparency requirements for automated decision-making systems of this kind.

Community Creative Uses

Less than a week after launch, the community has already produced a wealth of imaginative use cases:

• jevchat (Kyle Pena) turned Jev into a (very bad) chat model: "At each step, ask Jev a question: given the user's question and the response written so far, what is the next token?" One commenter described it as "a numeric Morty talking to Death Crystals."
• jev-leftpad (Fatih Kadir Akın) implemented the classic left-pad function using Jev, with the prompt "How many spaces need to be prepended to value to reach targetLength?" — using a choice question from "0 spaces needed" to "10 spaces needed."
• jev-2048 (Andy Gayton) used Jev to play the 2048 sliding tile puzzle game.

The Open-Source Replication Wave

A wave of projects attempting to replicate similar capabilities using open-weight models has also emerged around Jev. Kev is an interesting example, built on Qwen in 0.8B, 4B, and 9B sizes. Meanwhile, a benchmark called JevBench has already appeared, specifically designed to compare "Jev-class decision models."

Considering Jev has been out for less than a week, the activity it has generated is impressive. Whether this type of model will truly establish itself as an independent category remains to be seen, but it raises a question worth pondering: when all you need is a decision rather than a paragraph of text, can large models be made faster, cheaper, and more focused?

Qwen (Tongyi Qianwen) is a large language model series open-sourced by Alibaba Cloud, known for excellent performance on Chinese-English bilingual tasks and a permissive open-source license. It has become a popular base for fine-tuning and derivative research in the international open-source community. Kev's choice to build decision models ranging from 0.8B to 9B parameters on Qwen reflects both the competitive inference efficiency of Qwen's smaller variants and the open-source community's ongoing exploration of "reproducing the core capabilities of closed-source commercial models using existing open weights." The emergence of JevBench signals that this model paradigm already has the potential to form an independent evaluation track, and standardized benchmarks are crucial for driving this class of models toward practical adoption.