The new JEV model, developed by Typeface, represents a significant step forward in artificial intelligence focused on machine tasks rather than human interaction. Unlike traditional models that generate text for humans, JEV is trained to make decisions for code execution. This allows for high speed and low operational costs, making it a promising tool for developers.

Key Features and Benefits of JEV

JEV was introduced by Typeface, a company founded by Diego Almeida, one of the co-founders of ChatGPT. The model differs from existing ones by focusing on decision-making rather than text generation. This is achieved through the use of the "reinforced learning with calibrated decisions" approach, developed by the Typeface team over two years.

Key benefits of JEV:

• Speed: Request processing takes milliseconds.
• Cost: Operations are 20-200 times cheaper compared to existing models for similar tasks.
• Reliability: The model demonstrates high accuracy and reduces the likelihood of errors such as "hallucinations" or loss of confidence.
• Specialization: JEV is not designed for text or dialogue generation but focuses on performing specific tasks, such as categorization or determining the truthfulness of statements.

Application of JEV

Thanks to its characteristics, JEV finds application in various fields:

• Agent Development: Enhances the accuracy and efficiency of code agents and agents in production systems.
• Task Routing: Can be used as a "model router" to determine which model is best suited for a specific task.
• Ticket Triage: Allows for rapid classification and prioritization of incoming tickets, which is significantly faster and cheaper than using traditional large language models.
• Classification and Verification: Companies involved in large-scale data classification or verification can significantly reduce costs. For example, Verscell reported a 6x speed-up in processes after switching from Gemini 2.5 Flash to JEV.
• Game Agents: Demonstrates the possibility of using JEV to control game characters, for instance, in the game Starcraft, where the model helped win missions.
• Browser Request Handling: Testing by Open Code company showed high performance of JEV when used in a browser.
• Smart Home: JEV can process natural language commands to control smart home devices, for example, "turn on the living room light," in 185 milliseconds, which is significantly faster than traditional LLMs.

Request Types and Model Structure

When interacting with JEV, three types of requests are used:

1. Choice: The model selects one option from a provided list. For example, when classifying a user's request into "billing," "technical," or "sales" categories, JEV can determine that the request pertains to technical support, along with a confidence level.
2. Score: The model assigns a score on a given scale. For instance, a user's frustration level can be assessed on a scale of "calm," "polite," or "angry."
3. New: The model determines the probability of truthfulness or the presence of a specific property. For example, "Is this true?" or "Is this urgent?".

JEV operates on a principle different from autoregressive models. Instead of generating tokens sequentially, JEV processes input data (state, text, JSON object) and a set of questions in parallel. This allows for high speed and efficiency.

Comparison with Traditional Software and LLMs

Traditional software is deterministic, operating on an "if-then" principle. LLM-based agents, in contrast, are not fully deterministic, making them flexible but sometimes prone to errors. JEV occupies a niche in hybrid AI systems, combining elements of determinism with AI-driven decision-making.

The JEV model does not generate tokens or use an autoregressive loop. Instead, it receives a state (e.g., text, JSON) and a set of questions, then evaluates them in parallel, returning a structured JSON response. This enables its use for decision-making tasks such as determining fraud risk levels, assessing confidence, or selecting an appropriate tool.

Access and Testing

Access to JEV is provided via an API. There is also a playground for testing and exploring the model's capabilities. Demonstrations include:

• "Six Degrees of Separation": Determining the number of steps required to go from one Wikipedia page to another (e.g., from "DNA" to "Manny Pu̲e̲r̲t̲o̲ ̲R̲i̲c̲o̲"). JEV completed this task in 1.7 seconds.
• Smart Home Control: Processing natural language commands to control devices.
• Email Prioritization: Classifying and prioritizing hundreds of emails in seconds.

Conclusion

JEV is a specialized model focused on making fast and cheap decisions for code. Its ability to process requests in milliseconds and with minimal cost opens up new possibilities for creating more efficient and scalable AI applications, especially in the field of code agents and workflow automation.