TypeSafe has launched Jev, an AI model designed for classification and scoring tasks rather than text generation. Founded by former OpenAI researcher Diogo Almeida, the model processes input states in string or JSON format and returns structured, typed responses based on predefined questions. It operates without traditional conversational capabilities, aiming to eliminate hallucinations by restricting outputs to a set list.
The model supports three types of queries: Score for numerical ratings and Noul for probability assessments. Jev is optimized for speed and cost-efficiency, with latency ranging from 70 to 500 ms and a price of $0.042 per million input tokens, while output tokens are free. The company claims the model is significantly faster and cheaper than frontier LLMs, though these benchmarks were conducted internally by TypeSafe.
Jev features a context window of 64k tokens, though the effective limit for state and question combinations is 32k. Performance may degrade as the input state grows. The service is primarily trained on English, with limited support for other languages. Current operational limits are set at 100,000 tokens per second and 80 requests per minute, though these specifications have been subject to recent adjustments.
The model is intended for use cases such as ticket routing, content moderation, and guardrails for other LLMs, replacing traditional regex-based systems. Because the model relies on probability thresholds for decision-making, developers are advised to validate its calibration against their own datasets, especially for non-English languages, as internal documentation examples have shown variability in output consistency.
