vLLM Opens Multimodal Decision Models for AI Routing
vLLM Semantic Router released Decision 3.0 models from 0.6B to 26.09B parameters that return structured choices instead of generated prose.
The vLLM Semantic Router team has released Decision 3.0, a family of open-weight models designed to answer typed questions about text, JSON, images and video. Instead of generating an explanation, the models return choices, yes-or-no decisions or scores with probabilities.
A different role for models
The release includes several sizes, ranging from 0.6-billion-parameter versions to the 26.09-billion-parameter d3 model. The models are intended for routing requests between specialist systems, applying moderation rules, selecting tools and enforcing policy gates. Their Apache-2.0 licensing allows developers to inspect, adapt and deploy the weights under a relatively permissive open-source framework.
The model card describes d3 as a 26.09-billion-parameter multimodal model, including a 0.46-billion-parameter vision encoder. It is built on Qwen3.8-27B and was trained on AMD Instinct MI325X GPUs. The design removes the need to generate text and then parse that text into a downstream decision, potentially reducing latency and simplifying failure handling.
Why it matters
Decision 3.0 reflects a broader shift in AI infrastructure: not every important model needs to be a general-purpose chatbot. In production systems, a small, fast and inspectable classifier can determine whether a request should reach a costly frontier model, which tool an agent may call or whether content requires human review.
The limits are equally important. A calibrated probability does not prove that the chosen action will succeed, and benchmark scores do not establish reliability under changing policies or adversarial inputs. Decision 3.0 is therefore most consequential as an open building block for model orchestration, not as a replacement for general-purpose reasoning models.