Baseten releases joey-1 for decisions instead of prose

Baseten announced joey-1 on October 29 for software that needs a decision before its next step. Its announcement emphasizes frequent, repeated classifications and cases where each decision adds waiting time to a response. The company positions this as an alternative to asking a language model to write an answer that software must then interpret.

A routing example

The official API example supplies a customer message about a duplicate charge as input. A separate question asks which team should handle it, with billing and technical support defined as options. The request includes an API key. This demonstrates how to describe the choice; it is not a recorded customer result or a refund operation. Applications can use the returned probabilities to select an answer or apply confidence thresholds.

Question formats and limits

The model card distinguishes yes/no checks, named alternatives and ordered scoring levels. For scores, the levels must start with the lowest. Inputs can contain text and images. Its self-hosting instructions require a CUDA GPU and a pinned PyTorch runtime. Requests exceeding the trained length limits are rejected rather than shortened.

MarkTechPost’s coverage also describes the release. That report summarizes the provider’s materials; it does not establish an independent test of this workflow.