vLLM Semantic Router releases d3-ultra, a 27B decision model

The vLLM Semantic Router team released d3-ultra, an open multimodal decision model with 26.09B parameters under Apache-2.0. It does not generate text; instead it reads the input once and returns probabilities for every question, covering routing, classification, visual grounding and video checks. It supports 131,072 tokens of context and was tested on a single AMD Instinct MI325X.

• 26.09B parameters, including a 0.46B vision encoder
• 131,072-token context under the Apache-2.0 license
• Best score 97.8 on SR-Text, worst 54.6% on SR-Box
• 255 questions about one photo take 2.16 s on MI325X