TypeSafe has introduced Jev, a new AI model developed by former OpenAI engineer Diogo Almeida. The company claims the model is up to 194 times faster and 445 times cheaper than frontier LLMs. Unlike traditional chatbots, Jev is designed for probabilistic decision-making rather than generating prose.
The model utilizes Reinforcement Learning for Calibrated Decisions to provide structured, machine-readable outputs in JSON format. It processes individual requests in parallel without requiring global context or memory. Each response includes a confidence factor, allowing developers to integrate the model directly into automated logic workflows.
Jev operates with a context window capped at 64,000 tokens. The system is intended for specific tasks such as intent routing or data assessment, where developers use the provided confidence scores to trigger subsequent actions. It does not store user data or perform inferences beyond the provided state.
While Jev is optimized for speed and efficiency, TypeSafe notes that it is not intended to replace LLMs for open-ended reasoning tasks. The company suggests that developers may use Jev alongside traditional LLMs, employing Jev for rapid decision-making and LLMs for complex report generation or log analysis.
