New LLM Prioritizes Speed, Efficiency for Automated Workflows

TypeSafe AI has launched Jev, a new large language model (LLM) specifically engineered for machine-to-machine communication rather than human interaction. Founded by Diogo Almeida, a former OpenAI researcher and co-inventor of Reinforcement Learning from Human Feedback (RLHF), Jev aims to address the inefficiencies of general-purpose LLMs in automated workflows.

Addressing Token Waste and Latency Issues

General-purpose LLMs often generate verbose responses even for simple decisions, leading to unnecessary token consumption and higher costs. Jev tackles this by returning concise, categorical answers (like which tool to invoke or whether a request should be approved) with associated probability scores rather than lengthy natural language explanations.

According to TypeSafe AI, Jev can respond in as little as 70 milliseconds compared to several seconds for conventional LLMs. This speed advantage stems from its ability to generate decisions directly without sequentially processing text tokens.

Strategic Model Allocation

Technology consultant David Linthicum points out that enterprises currently use general-purpose LLMs universally, even when applications only require fast, bounded decisions—akin to using an enterprise service bus for a simple yes/no question. By handling routine decisions through probabilistic function calls, Jev enables organizations to reserve more expensive models for complex tasks requiring nuanced understanding.

Reduced Engineering Overhead

Jev’s structured outputs could also reduce the engineering effort needed to integrate LLMs into automated systems. Instead of building extensive prompt layers and validation mechanisms around free-form responses, developers can rely on Jev’s defined decision formats.

This approach allows for more explicit control flow in agentic applications—where application code determines what happens next based on the model’s response rather than embedding logic within prompts—making workflows easier to test and maintain.

Considerations for CIOs

While Jev offers significant advantages, Stephanie Walter of HyperFrame Research notes that developers must define questions, outputs, thresholds, and escalation paths in advance. Additionally, organizations need to validate the model’s accuracy on their own data and establish clear guidelines for interpreting probability scores.

Paul Chada of Doozer AI highlights a concern for regulated industries: Jev’s probabilistic nature may not provide sufficient auditability since it does not explain why a decision was made, only how confident the model was in its assessment.