The vLLM Semantic Router team has released Decision 3.0, a suite of open-source multimodal decision models under the Apache-2.0 license. Ranging in size from 0.8 billion to 27 billion parameters, the models process text, JSON, and image inputs to answer structured decision queries by returning calibrated probabilities rather than generating raw text.

Built on Qwen base architectures, all five sizes include vision encoders capable of processing up to 1.6-megapixel images. The system operates as a high-speed router or guardrail, allowing developers to execute classification, routing, and gating tasks in a single model forward pass without parsing unstructured text outputs.

According to internal evaluations published by the team, the 27B variant scored 64.1 on text evaluation tasks and 71.6 on vision tasks, outperforming competing open decision baselines. The release aims to streamline pipeline latency for production AI agent frameworks.

Why it matters

  • Enables faster agentic routing and guardrailing by replacing LLM text generation with direct probability scores.

  • Reduces infrastructure overhead using fine-tuned small models like d3-flash that outpace previous generation larger models.

Source: marktechpost.com