Jev: AI Models That Make Decisions Instead of Generating Text

Most modern AI applications rely on Large Language Models (LLMs), even when their primary task involves making relatively simple decisions.

For example:

• Determining whether a request should be approved.

• Selecting the appropriate tool for an AI agent.

• Classifying incoming messages or documents.

• Evaluating the risk level of an operation.

Traditional LLM-based workflows often involve generating text, parsing the response, validating its structure, and converting the output into application logic.

Jev introduces a different approach to this process.

What Is Jev?

On September 15, 2026, TypeSafe AI introduced Jev, its first model in a new category called System One Models.

Unlike conventional generative models, Jev is designed specifically for structured decision-making rather than producing natural-language responses.

Instead of generating paragraphs or JSON-formatted text, Jev processes input information and returns typed decisions accompanied by probabilities.

TypeSafe AI developed a training method called Reinforcement Learning for Calibrated Decisions (RLCD), designed to improve the reliability of the model's probability estimates.

How Jev Works

Jev supports three primary decision types:

1. Choice

Selects an option from a predefined set of possibilities. This can be used for classification, request routing, and tool selection.

2. Score

Evaluates information against an ordered rating scale, making it useful for prioritization, quality assessment, and risk scoring.

3. Noul

Returns a probability for a yes-or-no question, allowing applications to make decisions based on configurable confidence thresholds.

These capabilities allow developers to integrate probabilistic decision-making directly into software workflows without relying on autoregressive text generation.

Can Jev Run on an NVIDIA RTX 3090?

The original Jev model is currently available through TypeSafe AI's hosted API. Its model weights have not been publicly released, meaning it cannot currently be deployed locally on an RTX 3090.

However, an open-source project called OpenJev demonstrates similar decision-making techniques using the Qwen3.5-4B model.

OpenJev has been tested on an NVIDIA RTX 3090 with 24 GB of VRAM.

Performance Benchmark

In a published experiment involving 21 binary decisions on a single RTX 3090, OpenJev reported:

The direct probability approach was approximately 5.2 times faster in this specific benchmark.

However, the methods agreed on only 18 of the 21 decisions. These results demonstrate a potential performance advantage, not equivalent decision accuracy.

The benchmark measures OpenJev's implementation, not the original proprietary Jev model.

Practical Applications of Jev

1. AI Agents and Tool Selection

Decision-focused models can help AI agents select the appropriate tool, action, or workflow without generating lengthy intermediate responses.

2. Backend Microservices

Applications can use structured model outputs for semantic request routing, task classification, and intelligent workflow management.

3. Security and Fraud Detection

Probabilistic scoring can support suspicious activity detection, transaction risk assessment, and security event prioritization, with appropriate validation and human oversight.

4. Customer Support Automation

Incoming support requests can be classified by department, urgency, or issue type, improving ticket routing and prioritization.

5. Document Processing

Decision models can classify documents, identify relevant categories, and determine which processing pipeline should handle each document.

6. AI Orchestration

In multi-agent systems, specialized decision models can operate as routing components, determining which agent, service, or tool should execute the next operation.

Resources and References

Jev — TypeSafe AI: https://typesafe.ai/

Official Jev Announcement: https://typesafe.ai/blog/introducing-system-one-models-and-jev

OpenJev — GitHub: https://github.com/bonsai/openjev