Decision 3.0 Open Models Released by vLLM

The vLLM Semantic Router team has released Decision 3.0, a new family of multimodal decision models designed to process text, JSON, and images to answer typed questions. Available in five sizes ranging from 0.8B to 27B parameters under an Apache-2.0 license, Decision 3.0 aims to give developers a fast way to classify, route, and gate requests by returning calibrated probabilities instead of parsed text.

Built as fine-tunes of Qwen base models and equipped with integrated vision encoders, the models evaluate requests in a single call to deliver structured scores. Internal evaluations show that the 27B d3 model leads its text and vision tables on the Jev Decision Index 0.3.1, while smaller variants like the 9B d3-flash outperform the previous generation's 27B model. Latency tests conducted on an AMD Instinct MI325X GPU show median text processing speeds ranging from 21.7 milliseconds for the smallest model to 84 milliseconds for the largest variant.

This release matters because it provides developers with open-source, high-speed decision routing tools that streamline guardrails and classification tasks without the overhead of parsing traditional generative outputs.