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

The vLLM Semantic Router team has open-sourced d3-turbo under the Apache 2.0 licence on Hugging Face and ModelScope, alongside a free hosted System One API at decision.vllm-sr.ai and ROCm support in the vLLM SR Engine. Built as the 26.09B-parameter flagship of the Decision 3.0 family (including a 0.46B vision encoder), the model supports 32,768 tokens per question by default and 131,072 tokens when long_context=True is enabled.

Unlike generative language models that emit text token by token, d3-turbo encodes a text, JSON, image or video state once per request and evaluates up to 255 structured questions against that shared representation in a single pass, returning calibrated probabilities with zero output tokens. It supports five decision types: multiple choice (including images embedded inside individual questions or options), binary yes/no (noul), scalar scoring, single-object bounding boxes (box) and whole-video object tracking (track at 1 or 2 updates per second).

While the smaller d3-flash is faster for a single question on one AMD Instinct MI325X GPU (154 ms vs 302 ms), d3-turbo pulls ahead from four questions onward (328 ms vs 541 ms) and is 9.5 times faster at 64 questions (867 ms vs 8.26 s). On visual localisation and tracking evaluations, d3-turbo reaches 0.90 bounding-box accuracy at IoU 0.5 (0.96 confidence AUROC), 0.96 accuracy when images sit inside question options (+0.31 over d3) and 0.80 success AUC (0.97 presence F1) when tracking targets at 2 frames per second.