Baseten: joey-1

Joey 1 is a 2B-parameter open-weight decision model from Baseten, post-trained from Qwen3.5-2B. Rather than generating free-form text, it accepts structured context as input and returns calibrated probabilities for predefined choices. It handles classification, routing, and scoring in a single forward pass, and supports synchronous API calls for real-time use. Because the weights are open, it can also be downloaded and run in-house. This design makes it well suited to high-volume first-line decision making: content moderation pipelines, fraud detection, ticket triage and routing, spam and phishing detection, and prompt classification. A probability threshold of 0.85 is recommended when treating outputs as high-confidence decisions. It is intended for decision tasks rather than general text generation, so open-ended writing is out of scope. Joey 1 serves as the fast, cost-efficient high-volume tier of a hierarchical setup, where uncertain cases are escalated to larger models such as Elephant-1 or Mammoth-1. It is part of Baseten's model library and deployable as an API endpoint, backed by an inference stack with custom kernels, modern decoding techniques, and advanced caching, 99.99% uptime, fast cold starts, and scaling in any region and any cloud, including hybrid setups.