The model card lists an Apache-2.0 licence and describes a 26.09B-parameter dense model, including a 0.46B vision encoder.
d3-ultra accepts text or JSON with images and video. Rather than generating free-form text, it returns probabilities for constrained tasks including multiple-choice selection, yes/no answers, scores, bounding boxes and video tracks. The card says one input can be read once and used to answer many questions, with up to 131,072 context tokens per question.
The authors report median latency from 304 ms for one question to 2.16 seconds for 255 questions about one photo on a single AMD Instinct MI325X. A free API at decision.vllm-sr.ai and a Hugging Face Space are listed. The reported 87.2 SR-Text result and other benchmarks are author-reported; the source does not establish independent evaluations, deployment cost, or production performance in routing, guardrail, or moderation systems.