vLLM Semantic Router Releases d3-ultra: A 27B Open Multimodal Decision Model
vLLM Semantic Router has released d3-ultra, a 27-billion-parameter open-weight model built for multimodal decision-making. Unlike generative models focused on producing text or images, d3-ultra is designed around routing and classification - it takes in a combination of text, images, and video, processes them in a single pass, and returns a probability score for each candidate answer or route. The team reports a score of 87.2 on their internal SR-Text evaluation benchmark.
The model's multimodal input handling is a notable aspect of its design. Many decision and routing models operate on text alone, requiring separate preprocessing pipelines for visual content. By natively ingesting video and images alongside text, d3-ultra reduces the need for that kind of multi-stage architecture, which can introduce latency and error accumulation. This positions it as a practical option for applications where decisions need to be made on mixed-media content in a single inference step.
At 27 billion parameters, d3-ultra sits in a range that requires meaningful compute but is increasingly feasible for teams with access to modern GPU hardware or cloud inference services. The open-weight release means the model can be self-hosted, fine-tuned, or audited - an important consideration for organizations that cannot route sensitive content through third-party APIs. The vLLM ecosystem, known for its high-throughput inference server, provides a natural deployment environment for a model of this scale.
The SR-Text benchmark score of 87.2 gives a reference point for text-based decision accuracy, though independent evaluation across a wider range of tasks will be needed to understand where the model performs well and where it falls short. As multimodal AI systems become more common in production pipelines - handling documents, video feeds, and mixed inputs - specialized decision models like d3-ultra are likely to become a more prominent part of the tooling landscape.
