vLLM Semantic Router Releases d3-ultra: A 27B Open Multimodal Decision Model

vLLM Semantic Router has put out d3-ultra, a multimodal decision model under Apache-2.0 that is built to decide, not chat. It reads the input once and returns probabilities for answers instead of generating prose. For developers, that means one open-weight call can handle routing, classification, visual grounding, and video checks.

The model sits in the team’s Decision 3.0 family, which runs from 0.6B to 27B. Its dense size is listed at 26.09B parameters, including a 0.46B vision encoder, with 131,072 tokens of context per question. The config points to a Qwen3.5-style backbone with 64 layers, split between 48 linear-attention layers and 16 full-attention layers. The card does not name a base checkpoint.

Question handling is unusually flexible. A request can combine text or JSON with images or video, then ask choice, yes/no, score, box, or track questions. Images can even appear inside the questions and options themselves. If the input goes past 131,072 tokens, the model fails with max_length_exceeded rather than quietly chopping anything off.

On the team’s private tests, d3-ultra comes out on top in SR-Text, which contains 10,411 text requests and 73,495 questions pulled from public benchmarks plus 43 new task families. It scores 87.2 there, ahead of GLiDE V1.2 at 85.1, d1-ultra at 83.4, d3-turbo at 83.1, d3 at 81.4, and Jev 1.14 at 73.1. The model leads across the board except arts and human taste, where d1-ultra edges it by a little, 90.8 to 89.6.

The weaker spot is box grounding. On SR-Box, a set of 5,678 box questions across 4,984 images and 10 domains, d3-ultra scores 58.2% across all domains and 54.6% across 9. That trails d3-turbo, which posts 60.0% and 59.6%. The team says d3-turbo is stronger on natural images, while d3-ultra does better on charts. In latency testing on one AMD Instinct MI325X, a single photo question took 304 ms, and 255 questions took 2.16 seconds total.

There are a few ways to run it: Transformers with AutoModel.from_pretrained, the vLLM Semantic Router Engine, a free API, or an official Hugging Face Space. The model was tested on one AMD Instinct MI325X with BF16 weights, and no official quantized version is listed yet.