Vllm semantic router launches decision 3.0: a new family of multimodal decision models

The vLLM Semantic Router team has introduced Decision 3.0, a series of open-source multimodal decision models designed to process text, JSON, and images to answer typed questions. The models come in five sizes, ranging from 0.8B to 27B parameters, all under the Apache-2.0 license. They enable developers to classify, route, and gate requests efficiently without parsing generated text. Each model returns probabilities for questions, which can be Yes/No, Choice, or Score-based, providing calibrated scores for thresholding. Decision 3.0 models are fine-tuned versions of Qwen base models, with the 27B variant built on Qwen3.8-27B. All models include a vision encoder, with varying sizes from 0.10B to 0.46B. They accept multiple images up to 1.6 megapixels and process each question independently. The team reports performance metrics, with d3 scoring 64.1 on text tasks and 71.6 on vision tasks, outperforming Perplexity Decider v1.1 in several areas. However, performance varies across different benchmarks, with some models showing weaker results on specific vision tasks.