What is Decision 3.0? vLLM Semantic Router’s New Open Decision Models

vLLM Semantic Router has released Decision 3.0, a suite of five multimodal decision models spanning parameter counts from 0.8B to 27B. All models are released under the Apache 2.0 license, making them freely usable in commercial applications without restriction. The release marks a step toward purpose-built models that handle structured reasoning tasks rather than general text generation.

The core design principle behind Decision 3.0 is constraint - these models are built exclusively to answer decision-type questions rather than produce open-ended text. Supported query types include multiple-choice selection, binary yes/no judgments, and numerical scoring. This constrained output format means the models can return a usable decision in a single forward pass, which has practical implications for latency and cost in production pipelines.

Multimodal support is a notable aspect of the release. Decision 3.0 models accept both text and image inputs, meaning they can be used to make structured judgments about visual content as well as written content - useful in moderation, retrieval ranking, document classification, and other applied workflows where a model needs to evaluate rather than describe. The range of model sizes gives developers flexibility to trade off speed and resource use against accuracy depending on the deployment environment.

The broader context here is a growing interest in models that are specialized for routing and evaluation roles within larger AI systems. Rather than routing queries to a large general-purpose model for every decision, a small, fast decision model can handle classification and gating tasks much more efficiently. Decision 3.0 fits into this pattern, with the vLLM Semantic Router project positioning these models as components within inference pipelines rather than standalone assistants.