vLLM Semantic Router’s d3-ultra is a roughly 27B-parameter multimodal decision model that accepts text, JSON, images and video and returns option probabilities and scores. Unlike a chatbot generating extended prose, it reads the input once for multiple questions, including image localization and video tracking. Apache 2.0 weights and examples allow direct testing in classification and routing pipelines. Published latency figures are project measurements on specific GPU configurations; real-world accuracy, confidence calibration and hardware requirements still need evaluation on the intended data.
Back to Tech
d3-ultra returns decision probabilities, not prose
Multiple questions share one input pass for classification and localization