Jev

Overview. Jev is TypeSafe’s text-input model for typed decisions rather than generated prose. It returns decisions and probabilities through JSON, using reinforcement-learning post-training aimed at calibrated decisions. Developers can use it when an application needs a label, numeric score, or yes/no probability instead of a conversational answer.

Key strengths. Jev’s documented interface provides Choice, Score, and Noul response types for bounded decision tasks. These outputs let applications consume decisions and associated probabilities without parsing free-form prose. The available research contains no benchmark results establishing comparative accuracy or calibration performance.

When to use it. Use Jev for text classification when your application needs a selected label and probabilities over the available choices. Use its Score or Noul response types when a workflow needs a numeric assessment or yes/no probability. Integrate through TypeSafe’s System One API rather than assuming a chat-completions interface.

Limitations. The combined request budget is 64,000 tokens, but the state plus the longest individual question must fit within 32,000 tokens. Jev does not generate prose or visible reasoning traces and does not accept image, audio, or video inputs; English is its primary training language. The jev-latest alias can move between releases, so it should not be treated as a permanently pinned version.

Sovereignty note. TypeSafe AI, Inc. is a US company and serves Jev only from the United States, so requests are exposed to the US CLOUD Act. TypeSafe states that customer requests and responses are not used for training. Jev is not a Sovereign model, and a workspace with an EU-only or no-CLOUD-Act policy will not reach it.